WO2023150943A1 - Method for updating wireless channel model, and apparatus, device and storage medium - Google Patents
Method for updating wireless channel model, and apparatus, device and storage medium Download PDFInfo
- Publication number
- WO2023150943A1 WO2023150943A1 PCT/CN2022/075697 CN2022075697W WO2023150943A1 WO 2023150943 A1 WO2023150943 A1 WO 2023150943A1 CN 2022075697 W CN2022075697 W CN 2022075697W WO 2023150943 A1 WO2023150943 A1 WO 2023150943A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- wireless channel
- model
- discriminator
- channel model
- update
- Prior art date
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L25/00—Baseband systems
- H04L25/02—Details ; arrangements for supplying electrical power along data transmission lines
Definitions
- the present application relates to the communication field, and in particular to a method, device, equipment and storage medium for updating a wireless channel model.
- the neural network model In the application process of the neural network model, because the actual business data will change to varying degrees over time, the neural network model cannot be applied to new business data. Therefore, it is necessary to update the neural network model in time.
- the computer equipment sends the data output by the neural network model to the business module to execute the relevant business, and the business module indirectly feeds back the neural network according to the business performance index generated by the output data in the process of executing the business.
- the performance of the model when the business module feedbacks that the performance of the neural network model deteriorates, the neural network model is updated.
- the above method can only indirectly judge the reliability of the neural network model through the business performance indicators fed back by the business module, and cannot perceive the degradation of the neural network model in real time.
- Embodiments of the present application provide a method, device, device, and storage medium for updating a wireless channel model, which can sense the degradation of the neural network model in real time and update the neural network model in time.
- the technical scheme is as follows.
- a method for updating a wireless channel model is provided, the method is performed by a first communication device, and the method includes:
- first channel information through a first wireless channel model, where the first wireless channel model is a generator in a first generation confrontation network;
- the first discriminator When it is determined by the first discriminator based on the first channel information that the first wireless channel model satisfies the update condition, an update of the first wireless channel model is triggered, and the first discriminator generates for the first Discriminator in adversarial networks.
- a method for updating a wireless channel model is provided, the method is performed by a second communication device, and the method includes:
- the first wireless channel model is a generator in the first generative adversarial network
- the first discriminator is a discriminator in the first generative adversarial network.
- a first communication device includes:
- the first model module is used to obtain the first channel information through the first wireless channel model, and the first wireless channel model is a generator in the first generative confrontation network;
- a first updating module configured to trigger an update of the first wireless channel model when the first discriminator determines based on the first channel information that the first wireless channel model satisfies an update condition, the first discriminator
- the discriminator is the discriminator in the first generative confrontation network.
- a second communication device comprising:
- the second update module is configured to update the first wireless channel model when the first discriminator determines that the first wireless channel model satisfies the update condition based on the first channel information, the first channel information being the first channel information
- a wireless channel model is output, the first wireless channel model is a generator in the first generative adversarial network, and the first discriminator is a discriminator in the first generative adversarial network.
- a first communication device includes: a processor; wherein,
- the processor is configured to obtain first channel information through a first wireless channel model, where the first wireless channel model is a generator in a first generative confrontation network;
- the processor is configured to trigger an update of the first radio channel model when the first discriminator determines that the first radio channel model meets an update condition based on the first channel information, and the first discriminator
- the discriminator is the discriminator in the first generative confrontation network.
- a second communication device includes: a processor; wherein,
- the processor is configured to update the first wireless channel model when the first discriminator determines that the first wireless channel model satisfies an update condition based on the first channel information, the first channel information being the first channel information
- a wireless channel model is output, the first wireless channel model is a generator in the first generative adversarial network, and the first discriminator is a discriminator in the first generative adversarial network.
- a first communication device includes: a processor; a transceiver connected to the processor; a memory for storing executable instructions of the processor ;
- the processor is configured to load and execute the executable instructions to implement the method for updating the wireless channel model performed by the first communication device as described in the above aspect.
- a second communication device includes: a processor; a transceiver connected to the processor; a memory for storing executable instructions of the processor ;
- the processor is configured to load and execute the executable instructions to implement the method for updating the wireless channel model performed by the second communication device as described in the above aspect.
- a computer-readable storage medium wherein executable instructions are stored in the computer-readable storage medium, and the executable instructions are loaded and executed by the processor to implement the above aspects.
- a computer program product or computer program comprising computer instructions stored in a computer-readable storage medium readable by a processor of a computer device from a computer
- the storage medium reads the computer instruction, and the processor executes the computer instruction, so that the computer device executes the method for updating the wireless channel model described in the above aspect.
- a chip is provided, the chip includes a programmable logic circuit or a program, and the chip is used to implement the method for updating a wireless channel model as described in the above aspect.
- the wireless channel model When training the wireless channel model, the wireless channel model is used as a generator and a discriminator to form a generative confrontation network, and the confrontation training is performed to obtain a trained wireless channel model and a discriminator.
- the wireless channel model In the application stage, use the wireless channel model to perform wireless channel-related services, use the discriminator to evaluate the output data of the wireless channel model, and determine the quality of the wireless channel model according to the evaluation result of the discriminator.
- the wireless channel model is updated.
- This method can use the discriminator obtained from the training settlement training to evaluate the output data of the model in the application stage, so as to perceive the reliability of the model output data in real time, and update the model in time when the model output data is unreliable.
- Fig. 1 is a structural diagram of a neural network model provided by an exemplary embodiment of the present application
- Fig. 2 is a structural diagram of a neural network model provided by an exemplary embodiment of the present application.
- FIG. 3 is a schematic flowchart of a method for updating a wireless channel model provided in an exemplary embodiment of the present application
- FIG. 4 is a schematic flowchart of a method for updating a wireless channel model provided in an exemplary embodiment of the present application
- FIG. 5 is a schematic flowchart of a method for updating a wireless channel model provided in an exemplary embodiment of the present application
- FIG. 6 is a schematic flowchart of a method for updating a wireless channel model provided in an exemplary embodiment of the present application
- FIG. 7 is a schematic diagram of an implementation environment of a method for updating a wireless channel model provided by an exemplary embodiment of the present application.
- FIG. 8 is a flow chart of a method for updating a wireless channel model provided in an exemplary embodiment of the present application.
- FIG. 9 is a flow chart of a method for updating a wireless channel model provided in an exemplary embodiment of the present application.
- FIG. 10 is a schematic flowchart of a method for updating a wireless channel model provided in an exemplary embodiment of the present application.
- FIG. 11 is a flow chart of a method for updating a wireless channel model provided in an exemplary embodiment of the present application.
- FIG. 12 is a flow chart of a method for updating a wireless channel model provided in an exemplary embodiment of the present application.
- Fig. 13 is a schematic flowchart of an update method provided by an exemplary embodiment of the present application.
- FIG. 14 is a schematic flowchart of an update method provided by an exemplary embodiment of the present application.
- FIG. 15 is a flow chart of a method for updating a wireless channel model provided in an exemplary embodiment of the present application.
- Fig. 16 is a schematic flowchart of an update method provided by an exemplary embodiment of the present application.
- Fig. 17 is a schematic flowchart of an update method provided by an exemplary embodiment of the present application.
- FIG. 18 is a flow chart of a method for updating a wireless channel model provided in an exemplary embodiment of the present application.
- Fig. 19 is a schematic flowchart of an update method provided by an exemplary embodiment of the present application.
- Fig. 20 is a schematic flowchart of an update method provided by an exemplary embodiment of the present application.
- Fig. 21 is a structural block diagram of a first communication device provided by an exemplary embodiment of the present application.
- Fig. 22 is a structural block diagram of a second communication device provided by an exemplary embodiment of the present application.
- Fig. 23 is a schematic structural diagram of a communication device provided by an exemplary embodiment of the present application.
- Neural network Artificial Neural Networks (ANNs for short), also referred to as neural networks (NNs) or Connection Model, is a model that imitates the behavioral characteristics of animal neural networks and performs distributed parallel information processing. Algorithmic Mathematical Model.
- the basic structure of a simple neural network is shown in FIG. 1 , including: an input layer 101 , a hidden layer 102 and an output layer 103 .
- the input layer 101 is responsible for receiving data
- the hidden layer 102 processes the data
- the final result is generated in the output layer 103 .
- each node represents a processing unit, which can also be called a neuron.
- Multiple neurons form a layer of neural network, and multiple layers of information transmission and processing construct an overall neural network.
- neural network deep learning algorithms have been proposed in recent years, more hidden layers have been introduced, and feature learning is performed through layer-by-layer training of neural networks with multiple hidden layers, which greatly improves the learning of neural networks.
- processing capabilities and are widely used in pattern recognition, signal processing, optimization combination, anomaly detection, etc.
- CNN Convolutional Neural Networks
- its basic structure includes: an input layer 101 , multiple convolutional layers 104 , multiple pooling layers 105 , a fully connected layer 106 and an output layer 103 .
- the introduction of the convolutional layer 104 and the pooling layer 105 effectively controls the sharp increase of network parameters, limits the number of parameters and taps the characteristics of the local structure, improving the robustness of the algorithm.
- the neural network model obtained by training with a limited data set when performing actual business, because the business data will change to the same degree over time, the neural network model will not be suitable for new business data, and even the model will fail. .
- the model training module 201 uses the training data collected by the data collection module 204 to train a model, and deploys the trained model to the model reasoning module 202 for application.
- the model reasoning module 202 outputs output data according to the reasoning data in the data collection module 204, and sends the output data to the business application module 203 for executing the business.
- the model reasoning module 202 collects model performance feedback, evaluates whether the model needs to be updated based on historical performance, and feeds back the model performance to the model training module 201, so that the training module 201 updates the model, and sends the updated model to the model reasoning module 202.
- neural network-based CSI Channel State Information, channel state information
- channel estimation channel estimation
- CSI prediction channel prediction
- Channel estimation due to the complexity and time-varying nature of the wireless channel environment, the receiver's estimation and recovery of the wireless channel directly affects the recovery performance of the received data.
- the channel estimation and restoration process in the related art is shown in FIG. 4 .
- the transmitter will also transmit a series of pilot symbols known to the receiver (reference signal symbol 302), such as CSI-RS (Reference Signal, reference signal), DMRS ( Demodulation Reference Signal, demodulation reference signal), etc.
- the receiver receives the received data symbol 303 corresponding to the transmitted data symbol 301, and the received reference signal symbol 304 corresponding to the transmitted reference signal symbol 302, and performs channel estimation 311: the receiver performs channel estimation 311 according to the real pilot (transmitted reference signal symbol 302) and the received pilot Frequency (received reference signal symbol 304) uses methods such as the LS algorithm to estimate the channel information at the position of the reference signal. Then perform channel recovery 312 based on the estimated channel information: the receiver uses an interpolation algorithm to recover the channel information on the full time-frequency resource according to the channel information estimated at the pilot position (reference signal position), which is used for subsequent channel information Feedback or data recovery etc.
- CSI feedback In view of AI (Artificial Intelligence, artificial intelligence) technology, especially deep learning has achieved great success in computer vision, natural language processing, etc., the field of communication has begun to try to use deep learning to solve technical problems that are difficult to solve by traditional communication methods .
- the neural network architecture commonly used in deep learning is nonlinear and data-driven. It can extract features from the actual channel matrix data and restore the channel matrix information compressed and fed back by the UE (User Equipment) side as much as possible on the base station side. While restoring the channel information, it also provides a possibility for the UE side to reduce the CSI feedback overhead.
- the CSI feedback based on deep learning regards the channel information as the image to be compressed, uses the deep learning self-encoder to compress the channel information, and reconstructs the compressed channel image at the receiving end, which can preserve the channel information to a greater extent .
- the self-encoder includes an encoder and a decoder, wherein the encoder is deployed at the sending end (UE side), and the decoder is deployed at the receiving end (base station side/access network device side). That is, after the transmitting end obtains the channel information through channel estimation, the encoder compresses and encodes the channel information to obtain the encoding result, and the transmitting end sends the encoding result to the receiving end, that is, the compressed bit stream is fed back to the receiving end through the air interface feedback link. end.
- the receiving end receives the encoding result and uses the decoder to decode the encoding result to obtain channel information, that is, the decoder restores the channel information according to the feedback bit stream to obtain complete CSI feedback.
- the encoder can adopt the superposition of multiple fully connected layers
- the decoder can adopt the design of the convolutional layer and the residual structure.
- the internal network model structure of the encoder and decoder can be flexibly designed.
- CSI prediction Input the periodic CSI feedback information received by the receiving end into the neural network model, which can predict the CSI at non-measurement time, so as to obtain a more complete CSI.
- the receiving end receives periodic CSI feedback, and at the moment corresponding to the white box 306 and black box 307, there is no CSI feedback, where the black box 307 represents several moments in the future.
- the CSI received at the time corresponding to the shaded box 305 is fed back into the neural network model, so as to output the predicted CSI at the time corresponding to the black box 307 .
- GAN Generative Adversarial Networks
- Generative confrontation network is a training method of neural network model. Based on Figure 6, a confrontation training method of generative confrontation network is briefly described.
- the Generative Adversarial Network consists of a generator 308 and a discriminator 309 .
- the model parameters in the generator are first fixed, and the discriminator is trained: the input data is input into the generator to obtain generated data.
- the generated data output by the generator and the real data in the sample data set are respectively used as the input of the discriminator, which are respectively input into the discriminator to obtain the predicted label, and the loss value is calculated based on the predicted label and the actual label.
- the actual label is the generated label
- the data input to the discriminator is real data
- the actual label is the real label.
- the model parameters in the discriminator are fixed and the generator is trained.
- the generator and the discriminator are taken as a whole, the input data is input into the generator to obtain the generated data, the generated data is input into the discriminator to obtain the predicted label, and then the loss value is calculated based on the predicted label and the real label, based on the predicted label
- the error (loss value) with the real label is used to adjust the model parameters in the generator and train the generator, that is, to make the data output by the generator be recognized as real data by the discriminator, which means that the generated data generated by the generator is close to the real data .
- the model parameters of one network can be continuously fixed, another network can be trained, and the generator can be converged through iterative training, and finally a trained generator can be obtained.
- the trained generator can output generated data close to the real data based on the input data, while the discriminator fully learns the characteristics of the real data.
- the update mode of the wireless channel model of the neural network is mostly based on the model performance feedback at the inference side, and the performance index is quantified to determine whether the current neural network model needs to be updated. That is, model updating in the related art is driven by feedback data.
- the method in the related art is used to feed back the model performance data in real time, and the resource overhead of the air interface transmission is relatively large.
- the periodic feedback of performance data can reduce the overhead caused by model performance feedback, but sacrifices the real-time perception of model degradation.
- model performance feedback involves the interaction of multiple network elements, the feedback process is cumbersome, and it only indirectly reflects the performance of the neural network model, which belongs to relatively coarse-grained feedback.
- model performance feedback scheme requires continuous observation of performance feedback for a period of time to roughly quantify the degree of degradation of the neural network model and qualitatively determine whether the model is invalid.
- Such a feedback scheme does not guide how the model should be updated.
- the form of feedback is single and not diversified.
- the model may not be updated in time, let alone dynamically adjusted according to changes in data distribution, which makes the model fall into a dilemma where the update cannot keep up with the data changes, affecting the long-term continuous performance of the model.
- the neural network model is only responsible for predicting the output, and does not output reliability indicators. It is impossible to evaluate the reliability of the current output results in real time, resulting in The output data of the neural network model guides the decision-making of the business module with a certain degree of blindness.
- the above-mentioned model performance feedback scheme also belongs to post-event feedback and is a remedial measure. That is, the above-mentioned model update solution is a remedial update solution under the premise of sacrificing business performance.
- the performance of the model based on deep learning is strongly related to the data distribution. Due to the unstable wireless environment, the data distribution will inevitably be affected by factors such as time, environment, and system strategy. Therefore, neural network models cannot avoid failures.
- an embodiment of the present application provides a method for updating a wireless channel model, which can sense model degradation in a timely manner and actively update the model in a timely manner.
- the network architecture and business scenarios described in the embodiments of the present application are for more clearly illustrating the technical solutions of the embodiments of the present application, and do not constitute limitations on the technical solutions provided by the embodiments of the present application.
- the evolution of the technology and the emergence of new business scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.
- the network architecture 100 may include: a terminal device 10 and a network device, where the network device may include at least one of an access network device 20 and a core network device 30 .
- the terminal equipment 10 may refer to a user equipment (User Equipment, UE), an access terminal, a subscriber unit, a subscriber station, a mobile station, a mobile station, a remote station, a remote terminal, a mobile device, a wireless communication device, a user agent or a user device.
- UE User Equipment
- the terminal device 10 may also be a cellular phone, a cordless phone, a Session Initiation Protocol (Session Initiation Protocol, SIP) phone, a Wireless Local Loop (Wireless Local Loop, WLL) station, a Personal Digital Assistant (PDA) ), handheld devices with wireless communication functions, computing devices or other processing devices connected to wireless modems, vehicle-mounted devices, wearable devices, terminal devices in the fifth generation mobile communication system (5th Generation System, 5GS) or future evolution
- the terminal equipment in the Public Land Mobile Network (Public Land Mobile Network, PLMN), etc. is not limited in this embodiment of the present application.
- the devices mentioned above are collectively referred to as terminal devices.
- the number of terminal devices 10 is generally multiple, and one or more terminal devices 10 may be distributed in a cell managed by each access network device 20 .
- the access network device 20 is a device deployed in an access network to provide a wireless communication function for the terminal device 10 .
- the access network device 20 may include various forms of macro base stations, micro base stations, relay stations, access points, and so on.
- the names of devices with access network device functions may be different.
- they are called gNodeB or gNB.
- the name "access network equipment” may change.
- the above-mentioned devices that provide the wireless communication function for the terminal device 10 are collectively referred to as access network devices.
- a communication relationship can be established between the terminal device 10 and the core network device 30.
- the access network device 20 may be an Evolved Universal Terrestrial Radio Access Network (Evolved Universal Terrestrial Radio Access Network, EUTRAN) or one or more eNodeBs in EUTRAN;
- EUTRAN Evolved Universal Terrestrial Radio Access Network
- the access network device 20 may be a radio access network (Radio Access Network, RAN) or one or more gNBs in the RAN.
- RAN Radio Access Network
- the network device refers to the access network device 20, such as a base station.
- the core network device 30 is a device deployed in the core network.
- the functions of the core network device 30 are mainly to provide user connections, manage users, and carry out services, and provide an interface to external networks as a bearer network.
- the core network equipment in the 5G NR system can include access and mobility management function (Access and Mobility Management Function, AMF) network element, user plane function (User Plane Function, UPF) network element and session management function (Session Management Function) Function, SMF) network element, etc.
- AMF Access and Mobility Management Function
- UPF User Plane Function
- SMF Session Management Function
- the access network device 20 and the core network device 30 communicate with each other through a certain air interface technology, such as the NG interface in the 5G NR system.
- the access network device 20 and the terminal device 10 communicate with each other through a certain air interface technology, such as a Uu interface.
- the "5G NR system" in the embodiments of the present disclosure may also be called a 5G system or an NR system, but those skilled in the art can understand its meaning.
- the technical solution described in the embodiments of the present disclosure can be applied to the 5G NR system, and can also be applied to the subsequent evolution system of the 5G NR system.
- Fig. 8 shows a flowchart of a method for updating a wireless channel model provided by an exemplary embodiment of the present application. This embodiment is described by taking the method applied to a first communication device as an example.
- the first communication device may be the terminal 10 or the network device 20 shown in FIG. 1 .
- the method includes:
- Step 410 Obtain first channel information through a first wireless channel model, where the first wireless channel model is a generator in the first generative adversarial network.
- the first wireless channel model is a neural network model that is trained as a generator against a first discriminator.
- the method for updating the wireless channel model is applied to the wireless channel model as an example for illustration.
- This method can also be applied to other neural network models in the communication field, or other fields (for example, image processing, natural language processing)
- the neural network model in .
- the names of other neural network models can be used to replace the "wireless channel model" in the embodiment of the present application.
- step 410 is a step of applying the trained first wireless channel model.
- the first wireless channel model is used to obtain the first channel information based on the second channel information. Use the first channel information to perform wireless channel-related services, or guide decisions on wireless channel-related services.
- the first radio channel model and the second radio channel model are only used to distinguish the radio channel model before updating and the radio channel model after updating, and both may be referred to as radio channel models.
- the network structure of the wireless channel model may be arbitrary, which is not limited in this embodiment of the present application.
- obtaining the first channel information based on the second channel information through the first wireless channel model refers to: the second channel information is input data, the first channel information is output data, and the second channel information is input into the first wireless channel
- the first wireless channel model outputs the first channel information based on the input second channel information.
- the second channel information is channel information input into the wireless channel model.
- the first channel information is channel information output by the wireless channel model based on the second channel information.
- the second channel information and the first channel information are the same type of channel information, or different types of channel information.
- Channel information is channel-related data
- channel information may include at least one of the following parameters: transmitting antenna, receiving antenna, time delay, FDM (Frequency Division Multiplexing, frequency division Multiplexing) number of symbols, number of subcarriers, eigenvectors obtained after channel eigenvalue decomposition, etc.
- FDM Frequency Division Multiplexing, frequency division Multiplexing
- the first wireless channel model and the first discriminator form the first generative adversarial network, wherein the first wireless channel model is used as the generator in the first generative adversarial network, and the first Wireless channel model and first discriminator.
- Step 420 When the first discriminator determines that the first wireless channel model satisfies the update condition based on the first channel information, trigger an update of the first wireless channel model.
- the first discriminator is a discriminator in the first GAN.
- a first discriminator is used to determine whether the first wireless channel model is degraded. Use the first discriminator to evaluate the first channel information output by the first wireless channel model to obtain the first probability value. If the first probability value satisfies the update condition, it means that the first wireless channel model meets the update condition, and triggers to update the first wireless channel model. channel model.
- the first probability value (first evaluation result) is obtained by evaluating the first channel information by the first discriminator.
- the first probability value is a value of 0-1, or, the first probability value is 0 or 1.
- the first discriminator is trained to determine whether the input data is real data. Utilizing the first discriminator to adversarially train the generator (the first wireless channel model) enables the generator to output channel information close to real channel information.
- step 420 continue to apply the first discriminator to judge the reliability of the first channel information.
- the discriminator is only used for confrontation training with the generator in the training phase, and in the application phase of the model, the discriminator is discarded and only the generator is used to perform business.
- the discriminator is creatively used to evaluate the output result of the generator in the application stage, so that whether the generator is degraded can be observed in real time based on the evaluation result output by the discriminator.
- the first discriminator and the second discriminator are only used to distinguish the pre-update discriminator from the updated discriminator, both of which can be called discriminators.
- the first discriminator is a discriminator trained against the first wireless channel model
- the second discriminator is a discriminator used for confrontation training with the second wireless channel model, or the second discriminator is trained against the second wireless channel model.
- the channel model is trained against the discriminator.
- the discriminator is a binary classification model.
- the output of the discriminator is 0 or 1, where 0 represents one class and 1 represents the other.
- 0 represents non-real data and 1 represents real data.
- the output of the discriminator is a value from 0 to 1, where the closer the value is to 0, the closer the data input to the discriminator is to the unreal data, the closer the value is to 1, the closer the data input to the discriminator is to real data, and 0.5 means Indicates that it is impossible to judge whether the data input to the discriminator is real data or non-real data.
- the update condition is used to determine whether the first wireless channel model is degraded.
- the first channel information output by the first wireless channel model is mostly judged as unreal data by the first discriminator (that is, the probability value is 0 or the probability value is lower than 0.5)
- the first wireless channel model has deteriorated and cannot output relatively real channel information, and the first wireless channel model needs to be updated. That is, degradation means that the channel information output by the first wireless channel model loses authenticity/reliability.
- the first wireless channel model may not be able to accurately extract the features in the new second channel information, and cannot output more realistic channel information. At this time It is necessary to update the first wireless channel model so that it can relearn the features in the second channel information after the data layout changes, and output more realistic channel data.
- Updating the first wireless channel model refers to continuing to train the first wireless channel model to obtain the second wireless channel model. That is, new training samples are used to train the first wireless channel model to obtain the second wireless channel model.
- the training samples include the second channel information, that is, use the real channel information input to the first wireless channel model during the application process to train the first wireless channel model, so that it can learn the characteristics of the second channel information and adapt to data changes .
- the wireless channel model when training the wireless channel model, the wireless channel model is used as a generator and a discriminator to form an adversarial network, and the adversarial training is performed to obtain a trained wireless channel model and a discriminator.
- the wireless channel model In the application stage, use the wireless channel model to perform wireless channel-related services, use the discriminator to evaluate the output data of the wireless channel model, and determine the quality of the wireless channel model according to the evaluation result of the discriminator.
- the wireless channel model is updated.
- This method can use the discriminator obtained from the training settlement training to evaluate the output data of the model in the application stage, so as to perceive the reliability of the model output data in real time, and update the model in time when the model output data is unreliable.
- Fig. 9 shows a flowchart of a method for updating a wireless channel model provided by an exemplary embodiment of the present application. This embodiment is described by taking the method applied to a second communication device as an example.
- the second communication device may be the terminal 10 or the network device 20 shown in FIG. 1 .
- the method includes:
- Step 310 When the first discriminator determines that the first wireless channel model satisfies the update condition based on the first channel information, update the first wireless channel model, the first channel information is output by the first wireless channel model, and the first wireless channel model
- the channel model is a generator in the first generation adversarial network
- the first discriminator is a discriminator in the first generation adversarial network.
- the second communication device cooperates with the first communication device to perform the wireless channel service.
- the second communication device sends the second channel information to the first communication device, and the first communication device inputs the second channel information into the first wireless channel model to obtain the first channel information, and applies the first channel information to for wireless channel business decisions.
- the second channel information is input information when the first wireless channel model generates the first channel information.
- the first communication device may use the first discriminator to determine whether the first radio channel model needs to be updated, and the first communication device may update the first radio channel model.
- the first communication device uses the first discriminator to determine whether the first wireless channel model needs to be updated, and when it needs to be updated, send an update instruction to the second communication device, instructing the second communication device to update the first wireless channel model to obtain the first wireless channel model.
- Two wireless channel models, and the second wireless channel model is sent by the second communication device to the first communication device.
- the first wireless channel model includes two parts: an encoder and a decoder.
- the encoder and the decoder are respectively deployed in the first communication device and the second communication device.
- the encoder side can use the first discriminator to judge whether the first wireless channel model needs to be updated, and the encoder side can update the first wireless channel model, and send the updated second decoder and second discriminator to the decoder side.
- the decoder side may use the first discriminator to judge whether the first wireless channel model needs to be updated, and the decoder side to send an update instruction to the encoder side, instructing the encoder side to update the first wireless channel model, and obtain after the update
- the second decoder and the second discriminator are sent to the decoder side.
- the encoder side may use the first discriminator to determine whether the first wireless channel model needs to be updated, and the encoder side to send an update instruction to the decoder side, instructing the decoder side to update the first wireless channel model, and obtain after the update
- the second encoder, the second decoder and the second discriminator are sent to the encoder side.
- the decoder side uses the first discriminator to judge whether the first radio channel model needs to be updated, and the decoder side updates the first radio channel model, and sends the updated second encoder to the encoder side.
- the first wireless channel model when the first wireless channel model is only deployed on a single-side communication device, four cases of division of labor between the first communication device and the second communication device are enumerated.
- Case 1 The first wireless channel model is deployed on the side of the first communication device, and the first communication device determines that the model is degraded, and performs an update of the wireless channel model.
- the second communication device sends the second channel information to the first communication device, and the first communication device receives the second channel information sent by the second communication device.
- the first communication device obtains the first channel information based on the second channel information through the first wireless channel model.
- the first communication device updates the first wireless channel model when it is determined by the first discriminator based on the first channel information that the first wireless channel model satisfies an update condition.
- Case 2 The first wireless channel model is deployed on the side of the first communication device, and the first communication device determines that the model is degraded, and instructs the second communication device to update the wireless channel model.
- the second communication device sends the second channel information to the first communication device, and the first communication device receives the second channel information sent by the second communication device.
- the first communication device obtains the first channel information based on the second channel information through the first wireless channel model.
- the first communication device sends an update instruction to the second communication device when it is determined by the first discriminator based on the first channel information that the first wireless channel model satisfies the update condition.
- the second communication device receives the update instruction, updates the first wireless channel model to obtain the second wireless channel model, and sends the second wireless channel model and the second discriminator to the first communication device.
- the first communications device receives and deploys a second wireless channel model and a second discriminator.
- Case 3 the first wireless channel model is deployed on the side of the first communication device, and the second communication device determines that the model is degraded, and performs an update of the wireless channel model.
- the second communication device sends the second channel information to the first communication device, and the first communication device receives the second channel information sent by the second communication device.
- the first communication device obtains the first channel information based on the second channel information through the first wireless channel model.
- the second communication device also obtains the first channel information based on the second channel information through the first wireless channel model.
- the second communication device updates the first wireless channel model when it is determined by the first discriminator based on the first channel information that the first wireless channel model satisfies the update condition.
- the second communication device sends the second wireless channel model and the second discriminator to the first communication device.
- the first communications device receives and deploys a second wireless channel model and a second discriminator.
- the second communication device sends the second channel information to the first communication device, and the first communication device receives the second channel information sent by the second communication device.
- the first communication device obtains the first channel information based on the second channel information through the first wireless channel model.
- the second communication device also obtains the first channel information based on the second channel information through the first wireless channel model.
- the second communication device sends an update instruction to the first communication device when it is determined by the first discriminator based on the first channel information that the first wireless channel model satisfies the update condition.
- the first communication device receives the update instruction, updates the first wireless channel model to obtain the second wireless channel model, and sends the second wireless channel model and the second discriminator to the second communication device.
- the first wireless channel model includes the first encoder and the first decoder
- four cases of division of labor between the first communication device and the second communication device are enumerated.
- Case 5 the first decoder is deployed on the side of the first communication device, and the first encoder is deployed on the side of the second communication device.
- the first communication device determines that the model is degraded, and performs an update of the wireless channel model.
- the second communication device obtains the first encoding result based on the second channel information through the first encoder, the second communication device sends the first encoding result to the first communication device, and the first communication device receives the first encoding result sent by the second communication device .
- the first communication device obtains the first channel information based on the first encoding result through the first decoder.
- the first communication device updates the first wireless channel model to obtain the second wireless channel model when it is determined by the first discriminator based on the first channel information that the first wireless channel model satisfies the update condition.
- the first communications device sends the second encoder to the second communications device.
- the second communication device receives the second encoder sent by the first communication device.
- Case 6 The first decoder is deployed on the side of the first communication device, and the first encoder is deployed on the side of the second communication device.
- the first communication device determines that the discriminant model is degraded, and instructs the second communication device to update the wireless channel model.
- the second communication device obtains the first encoding result based on the second channel information through the first encoder, the second communication device sends the first encoding result to the first communication device, and the first communication device receives the first encoding result sent by the second communication device .
- the first communication device obtains the first channel information based on the first encoding result through the first decoder.
- the first communication device sends an update instruction to the second communication device when it is determined by the first discriminator based on the first channel information that the first wireless channel model satisfies the update condition.
- the second communication device receives the more reliable indication, and updates the first wireless channel model to obtain the second wireless channel model.
- the second communication device sends the second decoder and the second discriminator to the first communication device.
- the first communication device receives the second decoder and the second discriminator sent by the second communication device.
- Case 7 The first decoder is deployed on the side of the first communication device, and the first encoder is deployed on the side of the second communication device.
- the second communication device discriminates that the model is degraded, and performs an update of the wireless channel model.
- the second communication device obtains the first encoding result based on the second channel information through the first encoder, the second communication device sends the first encoding result to the first communication device, and the first communication device receives the first encoding result sent by the second communication device .
- the first communication device obtains the first channel information based on the first encoding result through the first decoder.
- the second communication device also obtains the first channel information based on the first encoding result through the first decoder.
- the second communication device updates the first wireless channel model to obtain the second wireless channel model when it is determined by the first discriminator based on the first channel information that the first wireless channel model satisfies the update condition.
- the second communication device sends the second decoder and the second discriminator to the first communication device.
- the first communication device receives the second decoder and the second discriminator sent by the second communication device.
- Case 8 The first decoder is deployed on the side of the first communication device, and the first encoder is deployed on the side of the second communication device.
- the second communication device determines that the model is degraded, and instructs the first communication device to perform an update of the wireless channel model.
- the second communication device obtains the first encoding result based on the second channel information through the first encoder, the second communication device sends the first encoding result to the first communication device, and the first communication device receives the first encoding result sent by the second communication device .
- the first communication device obtains the first channel information based on the first encoding result through the first decoder.
- the second communication device also obtains the first channel information based on the first encoding result through the first decoder.
- the second communication device sends an update instruction to the first communication device when it is determined by the first discriminator based on the first channel information that the first wireless channel model satisfies the update condition.
- the first communication device receives the more reliable indication, and updates the first wireless channel model to obtain the second wireless channel model.
- the first communication device sends the second encoder, the second decoder and the second discriminator to the second communication device.
- the second communication device receives the second encoder, the second decoder, and the second discriminator sent by the second communication
- the first communication device in case 5 to case 8 can be replaced by the second communication device, and at the same time the second communication device is replaced by the first communication device, that is, the first decoder is deployed on the side of the second communication device, and the first The encoder is deployed on the side of the first communication device. In this way, four situations can also be obtained, which will not be described in detail in this embodiment.
- the wireless channel model when training the wireless channel model, the wireless channel model is used as a generator and a discriminator to form an adversarial network, and the adversarial training is performed to obtain a trained wireless channel model and a discriminator.
- the wireless channel model In the application stage, use the wireless channel model to perform wireless channel-related services, use the discriminator to evaluate the output data of the wireless channel model, and determine the quality of the wireless channel model according to the evaluation result of the discriminator.
- the wireless channel model is updated.
- This method can use the discriminator obtained from the training settlement training to evaluate the output data of the model in the application stage, so as to perceive the reliability of the model output data in real time, and update the model in time when the model output data is unreliable.
- Fig. 10 shows a structure diagram of a method for updating a wireless channel model provided by an exemplary embodiment of the present application. This embodiment is described by taking this architecture applied to the terminal 10 or the network device 20 shown in FIG. 1 as an example.
- the architecture consists of five main working modules: data collection module 501 , offline joint training module 502 , online model reasoning module 503 , online discriminator reasoning module 504 , and business application module 505 .
- Data collection module 501 as a data platform, it implements data preprocessing such as data filtering and data structuring, and provides training data and inference data to the offline joint training module 502 and the online model reasoning module 503.
- Offline joint training module 502 Driven by the training data, jointly train the first wireless channel model and the first discriminator, using the training method of the generative adversarial network for reference. By alternately training the model parameters of the two models until the models converge, the model training is completed, the trained first wireless channel model is deployed to the online model reasoning module 503 , and the trained first discriminator is deployed to the online discriminator reasoning module 504 .
- the first discriminator plays a supervisory role on the output of the first wireless channel model, such as supervising whether the output of the first wireless channel model conforms to the distribution characteristics of the real data. If it is consistent, the first discriminator outputs 1, and if it does not conform, the first discriminator outputs 0. .
- the output (evaluation result) of the discriminator may also be a discriminant index between 0 and 1.
- Online model reasoning module 503 After receiving the first wireless channel model sent by the offline joint training module 502, the online model reasoning module 503 completes model deployment.
- the offline joint training module 502 starts to intervene in the service, inputs the reasoning data (second channel information) into the first wireless channel model, and outputs the module output (first channel information) required by the service application module 505 .
- the module output is provided to the business application module 505 to perform business decisions.
- Online discriminator reasoning module 504 After receiving the first discriminator sent by the offline joint training module 502, the online discriminator reasoning module 504 completes the deployment of the first discriminator.
- the online discriminator reasoning module 504 starts to intervene in the business, and inputs the module output output by the first wireless channel model into the first discriminator, and the first discriminator performs discriminant reasoning on the model output to obtain the discriminator output (evaluation result).
- the discriminator output is sent to the business application module 505 to assist in the execution of business decisions.
- the architecture may also include: discriminator model online training module: when the online discriminator reasoning module 504 starts to intervene in the business, the discriminator model online training module simultaneously pre-updates the model parameters of the first discriminator online, and after the cache update The resulting second discriminator.
- the second discriminator is not used to perform a discrimination task (not used to discriminate the first channel information output by the first wireless channel model).
- the second discriminator is only used to memorize the distribution of new reasoning data (second channel information), learn the characteristics of the new reasoning data distribution in real time, and always save the model parameters most suitable for the current data distribution.
- the discriminator model online training module sends the second discriminator to the offline joint training module 502, and drives the offline joint training module 502 to train the first wireless channel model based on the second discriminator to obtain a model that is more in line with the current data distribution.
- Business application module 505 receives the module output sent by the online model reasoning module 503, and receives the discriminator output output by the online discriminator reasoning module 504. If the output of the discriminator satisfies the preset condition, the service application module 505 will directly use the output of the first wireless channel model as a basis for service decision. If the output of the discriminator does not meet the preset condition, the service application module 505 will not use the output of the first wireless channel model. Exemplarily, in this case, other strategy processes (the strategy of not using the neural network model process) to execute the business.
- the embodiment of the present application exemplarily provides multiple update condition setting schemes, and the embodiment of the present application also provides multiple update methods.
- Fig. 11 shows a flowchart of a method for updating a wireless channel model provided by an exemplary embodiment of the present application. This embodiment is described by taking the method applied to a first communication device as an example, and the first communication device may be the terminal 10 or the network device 20 shown in FIG. 1 .
- the method includes:
- Step 401 Use training samples to confront training an initial wireless channel model and an initial discriminator to obtain a first wireless channel model and a first discriminator.
- the first wireless channel model and the first discriminator are obtained through training.
- the wireless channel model and the model structure of the discriminator are determined, and model parameters are initialized to obtain an initial wireless channel model and an initial discriminator.
- Aiming at the generative adversarial network composed of the initial wireless channel model and the initial discriminator, the initial wireless channel model and the initial discriminator are trained against the training samples to obtain the first wireless channel model and the first discriminator.
- step 401 may also be performed by the second communication device, and when step 401 is performed by the second communication device, the first communication device receives the first wireless channel model and the first discriminator sent by the second communication device.
- Step 410 Obtain first channel information through a first wireless channel model, where the first wireless channel model is a neural network model that is trained as a generator against a first discriminator.
- the first channel information is obtained based on the second channel information through the first wireless channel model.
- the first channel information in step 410 may refer to one channel data, or may refer to multiple channel data output by the first wireless channel model based on multiple second channel information.
- Step 420 Evaluate the first channel information by a first discriminator to obtain a first probability value.
- the first discriminator outputs a probability value according to an input channel data.
- the first probability value includes at least one probability value corresponding to each channel data in the first channel information.
- the probability value is used to indicate whether the first channel information is real data or non-real data.
- the probability value includes 0 or 1, where 0 indicates non-real data and 1 indicates real data.
- real data refers to data obtained through actual collection or other data acquisition methods.
- Non-real data simulated data/generated data refers to the data output by the wireless channel model.
- the first discriminator may include multiple discriminators, and multiple discriminators are used to determine the authenticity of the channel information (whether it is real channel data) from multiple dimensions. Then the first probability value includes evaluation results output by multiple discriminators based on the first channel information.
- Step 421 Perform a wireless channel service based on the first channel information and the first probability value.
- both the first channel information and the first probability value may be applied to perform services related to wireless channels.
- the radio channel model only uses the first channel information to execute service decision, while the method provided in the embodiment of the present application also provides the first probability value to assist in the execution of service decision.
- the first probability value can evaluate the reliability of the first channel information. Taking the first channel information as an example, if the first probability value is not real data, it means that the first channel information is unreliable, and the first channel information cannot be used to perform business decisions; if the first probability value is real data , it indicates that the first channel information is reliable, and the first channel information can be used to perform service decisions.
- the first probability value is a value from 0 to 1
- the first probability value may also be used as the weight coefficient of the first channel information, so that when the first channel information is applied to In actual business, it reflects the reliability of the first channel information.
- the discriminator before the channel information output by the first wireless channel model is applied to the channel service, the discriminator can be used to sense whether the first wireless channel model is degraded in real time, without the need to apply the channel information After the information is applied to the channel service, the service feeds back the performance index.
- the method can perceive the degradation of the first wireless channel model in time before applying the channel information, so as to update the wireless channel model in time and prevent the model degradation from affecting service processing.
- Step 430 In a case where it is determined based on the first probability value that the first wireless channel model satisfies the update condition, update the first wireless channel model to obtain a second wireless channel model.
- the determination of whether the update condition is satisfied may be performed in real time, after outputting a specified amount of channel information, or periodically. That is, after each piece of channel information is output by the first wireless channel model, it may be determined whether the update condition is satisfied based on the probability value of the channel information, so as to update the wireless channel model in real time. It may also be determined whether the update condition is satisfied based on the probability value of the specified amount of channel information after the first wireless channel model outputs the specified amount of channel information. It is also possible to periodically determine whether the update condition is met according to the probability value obtained in the period.
- the first channel information includes a plurality of channel information
- the first probability value includes a plurality of probability values corresponding to the plurality of channel information
- the discriminator is used to distinguish whether the input data is real channel data (real data) or non-real Channel data (not real data).
- the update condition is a criterion for evaluating whether the wireless channel model is degraded based on the first probability value.
- the setting of the update condition can be arbitrary, and the method of expressing the degree of degradation of the wireless channel model based on the first probability value can be used to realize The setting of the update condition, the embodiment of the present application does not limit the setting of the update condition.
- Update conditions include at least one of the following four conditions:
- Condition 1 The proportion of the probability values lower than the first threshold among the first probability values is higher than the preset value.
- the proportion of the probability value in the first probability value that is not real channel data (that is, the probability value is lower than the first threshold, for example, lower than 0.5) is higher than the preset value.
- the degradation degree of the wireless channel model is represented by the ratio of the non-real channel data, the higher the ratio, the higher the degradation degree of the wireless channel model; the lower the ratio, the better the wireless channel model.
- the proportion can be calculated once, that is, the first probability value includes a plurality of probability values recently output by the first discriminator, and the number of recent output by the first discriminator is calculated.
- a probability value represents the proportion of non-real channel data. If the proportion is high, it indicates that the model is degraded and needs to be updated.
- the proportion may also be calculated once after the first wireless channel model outputs a specified amount of channel information, that is, the first probability value includes the specified number of probability values. For example, whenever the first wireless channel model outputs 100 pieces of channel information, the proportion of non-real channel data is calculated based on 100 probability values of the 100 pieces of channel information.
- the proportion can also be calculated periodically, that is, the proportion can be calculated based on multiple probability values output by the first discriminator within the calculation time at intervals. That is, the first probability value includes a plurality of probability values output by the first discriminator in the first cycle, and the cycle duration of the first cycle is a preset duration. For example, every hour, based on 58 probability values of 58 channel information output by the first wireless channel model within 1 hour, the proportion of non-real channel data is calculated.
- Condition 2 x consecutive probability values among the first probability values are lower than the second threshold.
- the x consecutive probability values in the first probability value are not real channel data (that is, the x consecutive probability values are lower than the second threshold, for example, lower than 0.5), and x is a positive integer.
- Condition 2 can also be expressed as: the latest x probability values output by the first discriminator are not real channel data. Or, among the plurality of probability values, there are consecutive x probability values that are not real channel data, wherein the plurality of probability values are output by the first discriminator in the first cycle, or, the plurality of probability values are output by the first discriminator Outputs the specified number of probability values.
- the first probability value includes the latest x probability values output by the first discriminator. Whenever the first discriminator outputs a probability value, it is determined whether the x probability values recently output by the first discriminator are all unreal channel data. If they are all unreal channel data, the first probability value satisfies the update condition, and the first The radio channel model needs to be updated.
- the first probability value may also be N probability values output by the first discriminator in the first period, N is an integer greater than x, and it is determined whether there are consecutive x probability values in the first probability value. is non-true channel data.
- the first probability value may also be M probability values continuously output by the first discriminator, M is a preset value, M is an integer greater than x, and it is determined whether there are consecutive x probability values in the first probability value All are non-true channel data.
- Condition 3 The probability value distribution obtained according to the first probability value satisfies the first condition.
- the probability value may be a value from 0 to 1 (including decimals and integers)
- the first wireless channel may be determined based on parameters on the probability value distribution graph/curve according to multiple probability value distribution graphs/curves Whether the model is degraded. For example, the reliability of the channel information output by the first wireless channel model is determined according to the curvature, slope, rate of change, clustering, etc. in the probability value distribution diagram/curve, and then whether the first wireless channel model is degraded.
- the fact that the probability value distribution satisfies the first condition means that the probability value distribution shows that the channel data is not credible/the channel data is unreal channel data.
- Condition 4 The evaluation value calculated based on the first probability value reaches the third threshold.
- evaluation values for evaluating the degree of model degradation may also be calculated based on the first probability value, and whether the evaluation value reaches a third threshold is used to determine whether the update condition is satisfied.
- the above four conditions can be set as update conditions independently, or can be set as update conditions in any combination.
- Condition 1 and Condition 2 can be set together as an update condition. That is, when the proportion of the probability values lower than the first threshold value in the first probability value is higher than the preset value, and the consecutive x probability values in the first probability value are lower than the second threshold value, the first probability value is determined The value satisfies the update condition, triggering the update of the first wireless channel model.
- the first discriminator may further include multiple discriminators, for example, the first discriminator includes two discriminators. Different discriminators evaluate the authenticity of the first channel information from different dimensions.
- the first discriminator includes at least two discriminators, and the first probability value includes at least two sub-probability values respectively output by the at least two discriminators. Then, when it is determined based on at least one sub-probability value in the at least two sub-probability values that the first wireless channel model satisfies the update condition, the update of the first wireless channel model is triggered.
- the embodiment of this application provides two ways to update the model:
- Method 1 Obtain the second channel data, use the second channel data (positive sample) as a training sample, train and update the first discriminator in real time to obtain the second discriminator; determine that the first wireless channel model satisfies the update condition based on the first probability value Finally, the second discriminator and the first wireless channel model are used to form a generated confrontation network, and the first wireless channel model is trained against to obtain the second wireless channel model.
- the first discriminator After applying the first wireless channel model to obtain the first channel information, use the first discriminator to evaluate the first channel information to obtain the first probability value, and at the same time, use the second channel information to train the first discriminator to obtain the second discriminator. That is, the first discriminator is continuously trained with the second channel information, so that the second discriminator can always learn the latest data distribution characteristics of service data.
- the trained second discriminator will not be used to evaluate the first channel information.
- the first discriminator is still used to evaluate the first channel information, that is, the first The discriminator is used to evaluate the first channel information, and the second discriminator is used to update the first wireless channel model.
- the second discriminator is used to form a generation confrontation network with the first wireless channel model, and the first wireless channel model is trained against to obtain the second wireless channel model. It is equivalent to that the first discriminator is updated in real time, and the first wireless channel model is updated after it is determined that the first wireless channel model satisfies the update condition based on the first probability value, and is fixed when the first wireless channel model is updated.
- the model parameters of the second discriminator remain unchanged, that is, there is no need to retrain the second discriminator, and the first wireless channel model can be directly trained against the second discriminator.
- the first discriminator is iteratively trained using the second channel information to obtain a second discriminator, and the second discriminator is used to update the first wireless channel model.
- the way to train the first wireless channel model is: for the second generative confrontation network composed of the first wireless channel model and the second discriminator, keep the parameters of the second discriminator unchanged, and use the training samples to train the first wireless channel model A second wireless channel model is obtained.
- the real-time iterative training of the first discriminator will not only use the second channel information (positive samples), but also use the first channel information output by the first wireless channel model as a negative sample, and iteratively train the first discriminator in real time Get the second discriminator.
- Method 2 After determining that the first wireless channel model satisfies the update condition based on the first probability value, obtain the second channel information; for the generative confrontation network composed of the first discriminator and the first wireless channel model, combine the second channel information and the first A channel information is used as a training sample, and confrontation training is performed to obtain a second wireless channel model and a second discriminator.
- Method 2 does not need to train the first discriminator in real time, but only needs to conduct alternate confrontation training based on the first wireless channel model and the first discriminator after the update conditions are met, and obtain the trained second wireless channel model and the second discriminator. Discriminator.
- Step 440 Obtain third channel information through the second wireless channel model.
- Step 450 Evaluate the third channel information by the second discriminator to obtain a second probability value.
- Step 460 When it is determined based on the second probability value that the second radio channel model satisfies the update condition, update the second radio channel model.
- the method provided in this embodiment provides multiple methods for setting update conditions to ensure real-time model update.
- the method of updating the discriminator in real time can be adopted, so that the discriminator can learn the change of data distribution in real time.
- the wireless channel model may also be able to learn the changes in data distribution, and then output more realistic channel information. This method can feed back the change trend of data layout to the model while triggering the model update, thereby guiding the direction of model update and improving the efficiency of model update.
- Fig. 12 shows a flowchart of a method for updating a wireless channel model provided by an exemplary embodiment of the present application. This embodiment is described by taking the method applied to a first communication device as an example, and the first communication device may be the terminal 10 or the network device 20 shown in FIG. 1 .
- the method includes:
- Step 510 Obtain a first channel estimation result through the first channel estimation model, where the first channel estimation model is a neural network model obtained by adversarial training as a generator and a first discriminator.
- the method is executed by an online environment module of a terminal device or a network device.
- the first channel estimation model and the first discriminator are obtained by adversarial training in an offline environment.
- the second communication device may be a terminal device or a network device, the second communication device sends the second channel information to the first communication device, and the first communication device obtains the first channel estimation result based on the second channel information through the first channel estimation model.
- the second channel information may be a received reference signal and a real reference signal. Or, the second channel information may be estimated channel data.
- the second channel information is the data input into the first channel estimation model
- the first channel estimation result is the data output by the first channel estimation model
- the second channel information is input into the first channel estimation model
- the first channel estimation model A first channel estimation result is output based on the input second channel information.
- step 601 is performed to jointly train the first channel estimation model and the first discriminator.
- a channel estimation model 701 and a discriminator 702 form a generative confrontation network.
- the input of the channel estimation model 701 is the pilot data set (received reference signal + real reference signal), or the input is the estimated channel data (for example, the estimated channel data obtained in the channel estimation step shown in Fig.
- the output is a channel data set (channel estimation result)
- the NMSE function is used as a loss function
- the error (loss value) between the channel estimation result and the real channel data is measured
- the model parameters of the channel estimation model 501 are updated based on the loss value.
- the input of the first discriminator 702 is a channel data set (channel estimation result), wherein the channel estimation result output by the channel estimation model is marked 0, and the actual real channel data is marked 1.
- the discriminator can be understood as a binary classifier, and the loss function of the discriminator can be a cross-entropy loss function.
- the discriminator outputs a number between 0 and 1. If the discriminator cannot tell whether the input is the channel estimation result or the real channel data, then output 0.5.
- the training process of the first channel estimation model and the first discriminator can be: first, initialize the channel estimation model and the discriminator; Input the training sample into the channel estimation model to obtain the channel estimation result (mark 0); the model parameters of the channel estimation model are fixed Unchanged, use the channel estimation results (negative samples) and real channel data (positive samples) to train the discriminator, let the discriminator clearly distinguish the real channel data and channel estimation results (binary classification), and get the trained discriminator, At this time, the discriminator can distinguish between real channel data and channel estimation results; then, the model parameters of the channel estimation model are released, the model parameters of the discriminator are fixed, and the channel estimation model is trained with training samples until the model converges. At this time, the channel estimation model generates The channel estimation result is more realistic and conforms to the potential characteristics of real channel data.
- training steps of the above-mentioned generative adversarial network only select a relatively basic and commonly used training method, and other training methods of generative adversarial networks can also be used for adversarial training.
- step 602 is executed, and the offline environment module sends the first channel estimation model and the first discriminator to the online environment module.
- the online environment module deploys the first channel estimation model and the first discriminator.
- step 604 is executed, using the first channel estimation model and the first discriminator for online reasoning (outputting channel estimation results and evaluation results).
- Step 520 Evaluate the first channel estimation result by the first discriminator to obtain the first probability value.
- step 605 is executed, the first discriminator always learns new data distribution online, and continuously updates the model parameters of the first discriminator to obtain the second discriminator.
- Step 530 In a case where it is determined based on the first probability value that the first channel estimation model satisfies the update condition, update the first channel estimation model to obtain a second channel estimation model.
- the first update instruction also includes a second discriminator, the second discriminator is obtained by using the second channel information to train the first discriminator online in real time; the second channel estimation model is based on the second discriminator offline confrontation training of the first channel estimation model owned.
- the evaluation results output by the first discriminator can be used to directly perceive the degree of degradation of the current first channel estimation model.
- the negative discriminant results output by the statistical discriminator for example, the number of times the discriminator output is less than 0.8
- the probability of a negative discriminative result exceeds a certain threshold for example, 30%
- Step 607 triggers model update when the degree of degradation exceeds the budget threshold.
- step 609 the offline environment module uses the second discriminator to update the first channel estimation model to obtain a second channel estimation model.
- the channel estimation result generated by the second channel estimation model is more in line with the latent distribution characteristics of the latest real channel data.
- step 610 the offline environment module sends the second channel estimation model to the online environment module.
- the online environment module deletes the first discriminator, uses the second discriminator to replace the first discriminator, uses the second channel estimation model to replace the first channel estimation model, and continues to execute the reasoning process.
- the channel estimation model when training the channel estimation model, the channel estimation model is used as a generator and a discriminator to form an adversarial network, and adversarial training is performed to obtain a trained channel estimation model and a discriminator.
- the channel estimation model is used to perform channel estimation-related services, and the discriminator is used to evaluate the output data of the channel estimation model, and the quality of the channel estimation model is determined according to the evaluation result of the discriminator.
- the channel estimation model is updated.
- This method can use the discriminator obtained from the training settlement training to evaluate the output data of the model in the application stage, so as to perceive the reliability of the model output data in real time, and update the model in time when the model output data is unreliable.
- Fig. 15 shows a flowchart of a method for updating a wireless channel model provided by an exemplary embodiment of the present application. This embodiment is described by taking the method applied to the first communication device as an example, and the first communication device may be the network device 20 (access network device) shown in FIG. 1 .
- the method includes:
- Step 710 Obtain a first CSI recovery result by using the first CSI autoencoder model, which is a neural network model obtained by training as a generator against the first discriminator.
- the first CSI autoencoder model includes a first CSI encoder and a first CSI decoder.
- the first CSI self-encoding model is used to encode and compress the CSI feedback information on the terminal equipment side, send the compressed encoding result to the access network equipment side, and decode the recovered CSI feedback information on the access network equipment side.
- the first CSI autoencoder model is used to encode and restore CSI feedback information, that is, ideally, the input CSI feedback information and output CSI feedback information (first CSI restoration result) of the first CSI autoencoder model are the same.
- the second channel information is CSI feedback information/CSI data.
- the second channel information is data input to the first CSI auto-encoding model, and the first CSI restoration result is output data of the first CSI auto-encoding model.
- the second channel information is input to the first CSI encoder to obtain a first encoding result, and the first encoding result is input to the first CSI decoder to obtain a first CSI recovery result.
- Step 510 the first CSI decoder and the first discriminator sent by the terminal device are received, and the first CSI decoder and the first discriminator are obtained through training by the terminal device.
- the first encoding result sent by the terminal device is received, where the first encoding result is obtained by the terminal device encoding real CSI data (second channel information) by using a first CSI encoder.
- Step 510 includes: decoding the first coding result by a first CSI decoder to obtain a first CSI restoration result (first channel information).
- the CSI self-encoding model includes a CSI encoder 703 and a CSI decoder 704 .
- the CSI encoder is deployed on the terminal device side, and the CSI decoder is deployed on the access network device side.
- the CSI encoder encodes the input CSI data to obtain the first encoding result
- the terminal device sends the first encoding result to the access network device
- the access network device uses the CSI decoder to decode the first encoding result to obtain the first CSI recovery result .
- the discriminator is used to distinguish the CSI recovery results from real CSI data.
- Step 720 Evaluate the first CSI restoration result by the first discriminator to obtain the first probability value.
- Step 730 When it is determined based on the first probability value that the first CSI autoencoder model satisfies the update condition, update the first CSI autoencoder model to obtain a second CSI autoencoder model.
- the second CSI autoencoder model includes a second CSI encoder and a second CSI decoder.
- a second update instruction is sent to the terminal device, and the second update instruction is used to instruct the terminal device to update the first CSI autoencoder model and the first discriminator ; receiving the second CSI decoder and the second discriminator sent by the terminal device, where the second CSI decoder and the second discriminator are obtained by adversarial training of the first CSI autoencoder model and the first discriminator.
- step 801 the terminal device trains a first CSI autoencoder model and a first discriminator.
- step 802 the terminal device sends the first CSI decoder and the first discriminator to the access network device.
- Step 803 the access network device deploys the first CSI decoder and the first discriminator.
- Step 804 the terminal device collects and saves the CSI data within a recent period of time.
- Step 805 the access network device uses the first CSI decoder and the first discriminator to infer online and output the first CSI restoration result and the first probability value.
- step 806 the access network device calculates the degree of degradation based on the probability value output by the first discriminator.
- Step 807 when the degree of degradation exceeds the budget threshold, trigger model update.
- the first discriminator can make immediate feedback. Combined with the discrimination result of the first discriminator, the access network device can immediately determine that the CSI restoration result analyzed by the first CSI decoder is qualified CSI data. Provide a reference for the next step of decision-making. When factors such as environmental changes lead to changes in data distribution, the access network device can use the first discriminator to directly perceive the degree of degradation of the CSI self-encoding model, and trigger a model update. Step 808, the access network device sends a second update instruction to the terminal device.
- step 809 the terminal device uses the CSI data saved in step 804 to update the first CSI autoencoder model and the first discriminator to obtain the second CSI autoencoder model and the second discriminator.
- step 810 the terminal device sends the second CSI decoder and the second discriminator to the access network device.
- the method provided in this embodiment when training the CSI autoencoder model, uses the CSI autoencoder model as a generator and a discriminator to form an adversarial network, conducts confrontation training, and obtains the trained CSI autoencoder model and discriminator device.
- the CSI self-encoding model is used to perform channel estimation-related services
- the discriminator is used to evaluate the output data of the CSI self-encoding model
- the quality of the CSI self-encoding model is determined according to the evaluation result of the discriminator.
- the CSI autoencoder model is updated.
- This method can use the discriminator obtained from the training settlement training to evaluate the output data of the model in the application stage, so as to perceive the reliability of the model output data in real time, and update the model in time when the model output data is unreliable.
- Fig. 18 shows a flowchart of a method for updating a wireless channel model provided by an exemplary embodiment of the present application. This embodiment is described by taking the method applied to the first communication device as an example, and the first communication device may be the network device 20 (access network device) shown in FIG. 1 .
- the method includes:
- Step 910 Obtain a first CSI prediction result through the first CSI prediction autoencoder model, which is a neural network model obtained by training as a generator against the first discriminator.
- the first CSI predictive self-encoding model includes a first CSI predictive encoder and a first CSI predictive decoder.
- the first CSI predictive self-encoding model is used on the side of the terminal device to obtain a second encoding result by encoding the CSI predictive encoder based on the CSI of consecutive N historical periods, and send the second encoding result to the access network device
- the second encoding result is decoded by a CSI prediction decoder to obtain the first CSI prediction result.
- the first CSI prediction result is a predicted CSI sequence obtained based on N historical period CSIs. That is, referring to the illustration in FIG. 5 , the first CSI prediction self-encoding model is used to obtain a future time-frequency CSI sequence based on CSI prediction of N historical periods.
- the first CSI prediction result includes a CSI sequence.
- the second channel information includes N historical period CSIs. Input the second channel information into the first CSI prediction encoder to obtain the second encoding result, and input the second encoding result into the first CSI prediction decoder to obtain the first CSI prediction result (first channel information).
- the access network device receives the first CSI predictive coder and the first discriminator sent by the terminal device, and the first CSI predictive coder and the first discriminator are trained by the terminal device.
- the access network device receives the second encoding result sent by the terminal device, where the second encoding result is obtained by the terminal device by encoding the CSI sequence through the first CSI predictive encoder.
- Step 910 includes: decoding the second encoding result by a first CSI prediction decoder to obtain a first CSI prediction result.
- the CSI prediction self-encoding model includes a CSI prediction encoder 706 and a CSI prediction decoder 707 .
- the CSI predictive encoder is deployed on the terminal device side, and the CSI predictive decoder is deployed on the access network device side.
- the CSI predictive encoder encodes the input CSI for N consecutive historical periods to obtain the second encoding result
- the terminal device sends the second encoding result to the access network device
- the access network device uses the CSI predictive decoder to decode the second encoding result
- the spatiotemporal relationship discriminator 705 and the temporal relationship discriminator 709 are used to evaluate the authenticity of the CSI prediction results in the spatiotemporal and temporal dimensions, respectively.
- Step 920 Evaluate the first CSI prediction result by the first discriminator to obtain the first probability value.
- the first discriminator includes a first temporal relationship discriminator and a first spatiotemporal relationship discriminator
- the first probability value includes a first temporal probability value and a first spatiotemporal probability value
- the temporal relationship discriminator is used to evaluate the authenticity of the CSI prediction results from the time dimension of the CSI prediction results.
- the spatio-temporal relationship discriminator is used to evaluate the authenticity of the CSI prediction results from the spatial dimension of the CSI prediction results.
- the timing relationship discriminator is to learn the temporal correlation characteristics of the actual CSI sequence, and judge whether the temporal correlation characteristics of the predicted CSI sequence (CSI prediction result) output by the CSI prediction self-encoding model conform to the temporal correlation characteristics of the actual CSI sequence.
- the spatio-temporal relationship discriminator learns the spatial characteristics of CSI itself and pays attention to the extraction of spatial features, so as to judge whether the spatial distribution characteristics of CSI (CSI prediction results) output by the CSI prediction autoencoder model conform to the spatial distribution characteristics of actual CSI data.
- Step 930 When it is determined based on the first probability value that the first CSI predictive auto-encoding model satisfies the update condition, update the first CSI predictive auto-encoder model to obtain a second CSI predictive auto-encoder model.
- the second CSI predictive autoencoder model includes a second CSI predictive encoder and a second CSI predictive decoder.
- the first CSI predictive auto-encoding model is updated to obtain the second CSI predictive auto-encoder model.
- the access network device receives the second CSI predictive decoder and the second discriminator sent by the terminal device; wherein, the second CSI predictive decoder and the second discriminator are terminal devices, and the first timing probability value and the second discriminator In a case where at least one of the spatiotemporal probability values satisfies the update condition, the first CSI prediction obtained from the self-encoding model and the first discriminator is updated.
- step 901 the terminal device trains a first CSI prediction autoencoder model and a first discriminator.
- Step 902 the terminal device sends the first CSI prediction decoder and the first discriminator to the access network device.
- Step 903 the access network device deploys a first CSI prediction decoder and a first discriminator.
- Step 904 the terminal device learns online, continuously updates the model parameters of the first discriminator by using the CSI prediction results generated in the application process (reasoning process) and the real CSI sequence in the database to obtain the second discriminator.
- Step 905 the access network device uses the first CSI prediction decoder and the first discriminator to infer online and output the first CSI prediction result and the first probability value.
- Step 906 the terminal device calculates the degree of degradation based on the probability value output by the first discriminator.
- Step 907 when the degree of degradation exceeds the budget threshold, trigger model update.
- the first discriminator can make immediate feedback.
- the access network device can immediately determine whether the CSI prediction result analyzed by the CSI prediction decoder is reliable. Provide a reference for the next step of decision-making.
- the first discriminator can always learn new data distribution on the terminal device side and pre-update model parameters.
- the terminal device can directly perceive the degree of degradation of the CSI prediction self-encoding model by using the first initial discriminator, and trigger an update.
- the terminal device adversarially trains the first CSI predictive autoencoder model by using the second discriminator obtained by updating in real time to obtain the second CSI predictive autoencoder model.
- Step 908 the terminal device sends the second CSI prediction decoder and the second discriminator to the access network device.
- the method provided in this embodiment when training the CSI predictive autoencoder model, uses the CSI predictive autoencoder model as a generator and a discriminator to form an adversarial network, conducts confrontation training, and obtains a trained CSI predictive autoencoder Model and Discriminator.
- the application stage use the CSI predictive self-encoding model to perform channel estimation-related services, and use the discriminator to evaluate the output data of the CSI predictive self-encoding model, and determine the pros and cons of the CSI predictive self-encoding model according to the probability value of the discriminator.
- the CSI prediction self-encoding model is updated.
- This method can use the discriminator obtained from the training settlement training to evaluate the output data of the model in the application stage, so as to perceive the reliability of the model output data in real time, and update the model in time when the model output data is unreliable.
- Fig. 22 shows a block diagram of a first communication device provided by an exemplary embodiment of the present application, and the device includes:
- the first model module 21 is used to obtain the first channel information by the first wireless channel model, and the first wireless channel model is a generator in the first generated confrontation network;
- the first update module 22 is configured to trigger an update of the first wireless channel model when the first discriminator determines that the first wireless channel model satisfies an update condition based on the first channel information, and the first The discriminator is a discriminator in the first generation confrontation network.
- the first updating module 22 includes:
- the first sending submodule 24 is configured to send a first update instruction to a second communication device, where the first update instruction is used to trigger the second communication device to update the first wireless channel model to obtain a second wireless channel model;
- the first receiving submodule 23 is configured to receive the second wireless channel model sent by the second communication device.
- the first model module 21 is configured to obtain the first channel information based on the second channel information through the first wireless channel model;
- the device also includes:
- the first real-time training module 28 is used to use the second channel information as a training sample to train and update the first discriminator in real time to obtain a second discriminator;
- the first update instruction includes the second discriminator; the second discriminator is used to form a second generative adversarial network with the first wireless channel model to train and update the first wireless channel model.
- the first updating module 22 is configured to update the first wireless channel model to obtain a second wireless channel model.
- the first model module 21 is configured to obtain the first channel information based on the second channel information through the first wireless channel model;
- the device also includes:
- the first real-time training module 28 is used to use the second channel information as a training sample to train and update the first discriminator in real time to obtain a second discriminator;
- the first updating module 22 is configured to train and update the first wireless channel model to obtain the second wireless channel model based on the second generative confrontation network composed of the second discriminator and the first wireless channel model .
- the first updating module 22 is configured to train and update the first GAN based on the first GAN composed of the first discriminator and the first wireless channel model.
- a discriminator and the first radio channel model obtain a second discriminator and the second radio channel model.
- the device also includes:
- the first receiving module 27 is configured to receive the second channel information sent by the second communication device
- the first model module 21 is configured to obtain the first channel information based on the second channel information through the first wireless channel model.
- the first wireless channel model includes a first encoder and a first decoder, the first decoder is deployed on the side of the first communication device, and the first encoder is deployed On the side of the second communication device;
- the second wireless channel model includes a second encoder and a second decoder;
- the first receiving submodule 23 is configured to receive the second decoder and the second discriminator sent by the second communication device.
- the device also includes:
- the first receiving module 27 is configured to receive a first encoding result sent by the second communication device, where the first encoding result is obtained by the first encoder based on the second channel information;
- the first model module 21 is configured to use the first decoder to obtain the first channel information based on the first encoding result.
- the first wireless channel model includes a first encoder and a first decoder, the first encoder is deployed on the side of the first communication device, and the first decoder is deployed On the side of the second communication device, the first decoder is stored on the side of the first communication device; the second wireless channel model includes a second encoder and a second decoder;
- the first updating module 22 includes:
- the first sending sub-module 24 is configured to send the second decoder and the second discriminator to the second communication device.
- the first wireless channel model includes a first channel estimation model
- the first channel information includes a first channel estimation result
- the second channel information includes a reference signal.
- the first wireless channel model includes a first CSI self-encoding model
- the second wireless channel model includes a second CSI self-encoding model
- the first channel information includes a first CSI recovery result.
- the first communication device includes a terminal device, and the second communication device includes a network device;
- the first wireless channel model includes a first CSI predictive self-encoding model
- the second wireless channel model includes a second CSI predictive self-encoding model
- the first channel information includes a first CSI prediction result.
- the update module includes;
- the first evaluation submodule 26 is configured to evaluate the first channel information through the first discriminator to obtain a first probability value
- the first updating submodule 25 is configured to trigger updating of the first wireless channel model when it is determined based on the first probability value that the first wireless channel model satisfies an update condition.
- the first channel information includes a plurality of channel information
- the first probability value includes a plurality of probability values corresponding to the plurality of channel information one-to-one
- the update conditions include at least one of the following conditions:
- the proportion of probability values lower than the first threshold among the first probability values is higher than a preset value
- x consecutive probability values in the first probability value are lower than the second threshold, and x is a positive integer
- the probability value distribution obtained according to the first probability value satisfies the first condition
- the evaluation value calculated based on the first probability value reaches a third threshold.
- the first discriminator includes at least two sub-discriminators, and the first probability value includes at least two sub-probability values.
- the first update submodule 25 is configured to determine that the first wireless channel model satisfies an update condition based on at least one sub-probability value of the at least two sub-probability values , triggering an update of the first wireless channel model.
- FIG. 23 shows a block diagram of an apparatus for updating a wireless channel model provided by an exemplary embodiment of the present application.
- the apparatus can be realized as the above-mentioned second communication apparatus, and the apparatus includes:
- the second update module 32 is configured to update the first wireless channel model when the first discriminator determines that the first wireless channel model satisfies the update condition based on the first channel information, the first channel information being the
- the output of the first wireless channel model is that the first wireless channel model is a generator in the first generative adversarial network, and the first discriminator is a discriminator in the first generative adversarial network.
- the second updating module 32 includes:
- the second receiving submodule 33 is configured to receive the first update instruction sent by the first communication device, the first update instruction is determined by the first communication device based on the first channel information through the first discriminator Sent when the first wireless channel model satisfies an update condition;
- the second updating submodule 34 is configured to update the first wireless channel model to obtain a second wireless channel model
- the second sending submodule 35 is configured to send the second wireless channel model to the first communication device.
- the first update instruction includes a second discriminator
- the second discriminator is obtained by using the second channel information to train and update the first discriminator in real time, and the second channel The information is input information when the first wireless channel model generates the first channel information
- the second update submodule 34 is configured to train and update the first wireless channel model to obtain the second wireless channel model in the second generative confrontation network composed of the second discriminator and the first wireless channel model. channel model.
- the device also includes:
- the second sending module 36 is configured to send second channel information to the first communication device, where the second channel information is input information when the first wireless channel model generates the first channel information.
- the first wireless channel model includes a first encoder and a first decoder, the first decoder is deployed on the side of the first communication device, and the first encoder is deployed On the side of the second communication device;
- the second wireless channel model includes a second encoder and a second decoder;
- the second sending submodule 35 is configured to send the second decoder and the second discriminator to the first communication device.
- the device also includes:
- the second model module 31 is configured to use the first encoder to obtain a first encoding result based on second channel information, and the second channel information is an input when the first wireless channel model generates the first channel information information;
- the second sending module 36 is configured to send the first encoding result to the first communication device, and the first channel information is based on the first encoding result by the first communication device through the first decoder owned.
- the first wireless channel model includes a first channel estimation model
- the first channel information includes a first channel estimation result
- the second channel information includes a reference signal.
- the first wireless channel model includes a first CSI self-encoding model
- the second wireless channel model includes a second CSI self-encoding model
- the first channel information includes a first CSI recovery result.
- the second update module 32 is configured to update the first wireless channel model when it is determined based on the first probability value that the first wireless channel model satisfies the update condition, and the first A probability value is obtained by the first discriminator based on the first channel information.
- the first channel information includes a plurality of channel information
- the first probability value includes a plurality of probability values corresponding to the plurality of channel information one-to-one
- the update conditions include at least one of the following conditions:
- the proportion of probability values lower than the first threshold among the first probability values is higher than a preset value
- x consecutive probability values in the first probability value are lower than the second threshold, and x is a positive integer
- the probability value distribution obtained according to the first probability value satisfies the first condition
- the evaluation value calculated based on the first probability value reaches a third threshold.
- the first discriminator includes at least two sub-discriminators, and the first probability value includes at least two sub-probability values.
- the second update module 32 is configured to determine that the first wireless channel model satisfies an update condition based on at least one sub-probability value of the at least two sub-probability values, Updating the first wireless channel model.
- FIG. 23 shows a schematic structural diagram of a communication device (terminal or network device) provided by an exemplary embodiment of the present application.
- the communication device includes: a processor 1001 , a receiver 1002 , a transmitter 1003 , and a memory 1004 .
- the processor 1001 includes one or more processing cores, and the processor 1001 executes various functional applications and information processing by running software programs and modules.
- the receiver 1002 and the transmitter 1003 can be realized as a communication component, and the communication component can be a communication chip.
- the memory 1004 is connected to the processor 1001 .
- the memory 1004 may be used to store at least one instruction, and the processor 1001 is used to execute the at least one instruction, so as to implement various steps in the foregoing method embodiments.
- volatile or non-volatile storage devices include but not limited to: magnetic disk or optical disk, electrically erasable and programmable Electrically-Erasable Programmable Read Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Static Random Access Memory (SRAM), Read-Only Memory (Read-Only Memory, ROM), magnetic memory, flash memory, programmable read-only memory (Programmable Read-Only Memory, PROM).
- EEPROM Electrically-Erasable Programmable Read Only Memory
- EPROM Erasable Programmable Read Only Memory
- SRAM Static Random Access Memory
- Read-Only Memory Read-Only Memory
- PROM Programmable Read-Only Memory
- a computer-readable storage medium stores at least one instruction, at least one program, a code set or an instruction set, the at least one instruction, the At least one program, the code set or the instruction set is loaded and executed by the processor to implement the method for updating the wireless channel model performed by the terminal or network device provided in the above method embodiments.
- a computer program product or computer program comprising computer instructions, the computer instructions are stored in a computer-readable storage medium, the processor of the communication device can read from the computer The computer instruction is read by reading the storage medium, and the processor executes the computer instruction, so that the communication device executes the method for updating the wireless channel model performed by the terminal or network device described in the above aspect.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Molecular Biology (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Power Engineering (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Mobile Radio Communication Systems (AREA)
Abstract
A method for updating a wireless channel model, and an apparatus, a device and a storage medium, which relate to the field of communications. The method comprises: obtaining first channel information by means of a first wireless channel model, wherein the first wireless channel model is a generator in a first generative adversarial network (410); and when it is determined, by means of a first discriminator and on the basis of the first channel information, that the first wireless channel model meets an update condition, triggering the updating of the first wireless channel model, wherein the first discriminator is a discriminator in the first generative adversarial network (420). By means of the method, the degradation of a neural network model can be sensed in real time, and the neural network model can be updated in a timely manner.
Description
本申请涉及通信领域,特别涉及一种无线信道模型的更新方法、装置、设备及存储介质。The present application relates to the communication field, and in particular to a method, device, equipment and storage medium for updating a wireless channel model.
在神经网络模型的应用过程中,由于实际业务数据随着时间的推移会存在不同程度的变化,神经网络模型无法适用于新的业务数据。因此,需要对神经网络模型做及时更新。In the application process of the neural network model, because the actual business data will change to varying degrees over time, the neural network model cannot be applied to new business data. Therefore, it is necessary to update the neural network model in time.
相关技术中,在神经网络模型的应用过程中,计算机设备将神经网络模型输出的数据发送至业务模块执行相关业务,业务模块根据输出数据在执行业务过程中产生的业务性能指标来间接反馈神经网络模型的性能,在业务模块反馈神经网络模型性能变差时,对神经网络模型进行更新。In the related technology, during the application process of the neural network model, the computer equipment sends the data output by the neural network model to the business module to execute the relevant business, and the business module indirectly feeds back the neural network according to the business performance index generated by the output data in the process of executing the business. The performance of the model, when the business module feedbacks that the performance of the neural network model deteriorates, the neural network model is updated.
上述方法只能通过业务模块反馈的业务性能指标来间接评判神经网络模型的可靠性,无法实时感知神经网络模型的劣化。The above method can only indirectly judge the reliability of the neural network model through the business performance indicators fed back by the business module, and cannot perceive the degradation of the neural network model in real time.
发明内容Contents of the invention
本申请实施例提供了一种无线信道模型的更新方法、装置、设备及存储介质,可以实时感知神经网络模型的劣化并及时更新神经网络模型。所述技术方案如下。Embodiments of the present application provide a method, device, device, and storage medium for updating a wireless channel model, which can sense the degradation of the neural network model in real time and update the neural network model in time. The technical scheme is as follows.
根据本申请的一个方面,提供了一种无线信道模型的更新方法,所述方法由第一通信设备执行,所述方法包括:According to one aspect of the present application, a method for updating a wireless channel model is provided, the method is performed by a first communication device, and the method includes:
通过第一无线信道模型得到第一信道信息,所述第一无线信道模型为第一生成对抗网络中的生成器;Obtaining first channel information through a first wireless channel model, where the first wireless channel model is a generator in a first generation confrontation network;
在通过第一判别器基于所述第一信道信息确定所述第一无线信道模型满足更新条件的情况下,触发更新所述第一无线信道模型,所述第一判别器为所述第一生成对抗网络中的判别器。When it is determined by the first discriminator based on the first channel information that the first wireless channel model satisfies the update condition, an update of the first wireless channel model is triggered, and the first discriminator generates for the first Discriminator in adversarial networks.
根据本申请的一个方面,提供了一种无线信道模型的更新方法,所述方法由第二通信设备执行,所述方法包括:According to one aspect of the present application, a method for updating a wireless channel model is provided, the method is performed by a second communication device, and the method includes:
在通过第一判别器基于第一信道信息确定第一无线信道模型满足更新条件的情况下,更新所述第一无线信道模型,所述第一信道信息是所述第一无线信道模型输出的,所述第一无线信道模型为第一生成对抗网络中的生成器,所述第一判别器为所述第一生成对抗网络中的判别器。updating the first wireless channel model when it is determined by the first discriminator based on the first channel information that the first wireless channel model satisfies the update condition, the first channel information is output by the first wireless channel model, The first wireless channel model is a generator in the first generative adversarial network, and the first discriminator is a discriminator in the first generative adversarial network.
根据本申请的一个方面,提供了一种第一通信装置,所述装置包括:According to one aspect of the present application, a first communication device is provided, and the device includes:
第一模型模块,用于通过第一无线信道模型得到第一信道信息,所述第一无线信道模型为第一生成对抗网络中的生成器;The first model module is used to obtain the first channel information through the first wireless channel model, and the first wireless channel model is a generator in the first generative confrontation network;
第一更新模块,用于在通过第一判别器基于所述第一信道信息确定所述第一无线信道模型满足更新条件的情况下,触发更新所述第一无线信道模型,所述第一判别器为所述第一生成对抗网络中的判别器。A first updating module, configured to trigger an update of the first wireless channel model when the first discriminator determines based on the first channel information that the first wireless channel model satisfies an update condition, the first discriminator The discriminator is the discriminator in the first generative confrontation network.
根据本申请的一个方面,提供了一种第二通信装置,所述装置包括:According to one aspect of the present application, a second communication device is provided, the device comprising:
第二更新模块,用于在通过第一判别器基于第一信道信息确定第一无线信道模型满足更新条件的情况下,更新所述第一无线信道模型,所述第一信道信息是所述第一无线信道模型输出的,所述第一无线信道模型为第一生成对抗网络中的生成器,所述第一判别器为所述第一生成对抗网络中的判别器。The second update module is configured to update the first wireless channel model when the first discriminator determines that the first wireless channel model satisfies the update condition based on the first channel information, the first channel information being the first channel information A wireless channel model is output, the first wireless channel model is a generator in the first generative adversarial network, and the first discriminator is a discriminator in the first generative adversarial network.
根据本申请的一个方面,提供了一种第一通信设备,所述第一通信设备包括:处理器;其中,According to one aspect of the present application, a first communication device is provided, and the first communication device includes: a processor; wherein,
所述处理器,用于通过第一无线信道模型得到第一信道信息,所述第一无线信道模型为第一生成对抗网络中的生成器;The processor is configured to obtain first channel information through a first wireless channel model, where the first wireless channel model is a generator in a first generative confrontation network;
所述处理器,用于在通过第一判别器基于所述第一信道信息确定所述第一无线信道模型满足更新条件的情况下,触发更新所述第一无线信道模型,所述第一判别器为所述第一生成对抗网络中的判别器。The processor is configured to trigger an update of the first radio channel model when the first discriminator determines that the first radio channel model meets an update condition based on the first channel information, and the first discriminator The discriminator is the discriminator in the first generative confrontation network.
根据本申请的一个方面,提供了一种第二通信设备,所述第二通信设备包括:处理器;其中,According to one aspect of the present application, a second communication device is provided, and the second communication device includes: a processor; wherein,
所述处理器,用于在通过第一判别器基于第一信道信息确定第一无线信道模型满足更新条件的情况下,更新所述第一无线信道模型,所述第一信道信息是所述第一无线信道模型输出的,所述第一无线信道模型为第一生成对抗网络中的生成器,所述第一判别器为所述第一生成对抗网络中的判别器。The processor is configured to update the first wireless channel model when the first discriminator determines that the first wireless channel model satisfies an update condition based on the first channel information, the first channel information being the first channel information A wireless channel model is output, the first wireless channel model is a generator in the first generative adversarial network, and the first discriminator is a discriminator in the first generative adversarial network.
根据本申请的一个方面,提供了一种第一通信设备,所述第一通信设备包括:处理器;与所述处理器相连的收发器;用于存储所述处理器的可执行指令的存储器;其中,所述处理器被配置为加载并执行所述可执行指令以实现如上述方面所述的由第一通信设备执行的无线信道模型的更新方法。According to one aspect of the present application, a first communication device is provided, and the first communication device includes: a processor; a transceiver connected to the processor; a memory for storing executable instructions of the processor ; Wherein, the processor is configured to load and execute the executable instructions to implement the method for updating the wireless channel model performed by the first communication device as described in the above aspect.
根据本申请的一个方面,提供了一种第二通信设备,所述第二通信设备包括:处理器;与所述处理器相连的收发器;用于存储所述处理器的可执行指令的存储器;其中,所述处理器被配置为加载并执行所述可执行指令以实现如上述方面所述的由第二通信设备执行的无线信道模型的更新方法。According to one aspect of the present application, a second communication device is provided, and the second communication device includes: a processor; a transceiver connected to the processor; a memory for storing executable instructions of the processor ; Wherein, the processor is configured to load and execute the executable instructions to implement the method for updating the wireless channel model performed by the second communication device as described in the above aspect.
根据本申请的一个方面,提供了一种计算机可读存储介质,所述计算机可读存储介质中存储有可执行指令,所述可执行指令由所述处理器加载并执行以实现如上述方面所述的无线信道模型的更新方法。According to one aspect of the present application, a computer-readable storage medium is provided, wherein executable instructions are stored in the computer-readable storage medium, and the executable instructions are loaded and executed by the processor to implement the above aspects. The update method of the wireless channel model described above.
根据本申请的一个方面,提供了一种计算机程序产品或计算机程序,该计算机程序产品或计算机程序包括计算机指令,该计算机指令存储在计算机可读存储介质中,计算机设备的处理器从计算机可读存储介质读取该计算机指令,处理器执行该计算机指令,使得该计算机设备执行上述方面所述的无线信道模型的更新方法。According to one aspect of the present application, a computer program product or computer program is provided, the computer program product or computer program comprising computer instructions stored in a computer-readable storage medium readable by a processor of a computer device from a computer The storage medium reads the computer instruction, and the processor executes the computer instruction, so that the computer device executes the method for updating the wireless channel model described in the above aspect.
根据本申请的一个方面,提供了一种芯片,所述芯片包括可编程逻辑电路或程序,所述芯片用于实现如上述方面所述的无线信道模型的更新方法。According to one aspect of the present application, a chip is provided, the chip includes a programmable logic circuit or a program, and the chip is used to implement the method for updating a wireless channel model as described in the above aspect.
本申请实施例提供的技术方案至少包括如下有益效果:The technical solutions provided by the embodiments of the present application at least include the following beneficial effects:
在训练无线信道模型时,将无线信道模型作为生成器与判别器组成生成对抗网络,进行对抗训练,得到训练好的无线信道模型和判别器。在应用阶段,使用无线信道模型执行无线信道相关的业务,并使用判别器评价无线信道模型输出的数据,根据判别器的评价结果来确定无线信道模型的优劣。当根据评价结果确定无线信道模型劣化时,对无线信道模型进行更新。该方法可以利用训练结算训练得到的判别器,在应用阶段对模型的输出数据进行评价,从而实时感知模型输出数据的可靠性,在模型输出数据不可靠时及时对模型进行更新。When training the wireless channel model, the wireless channel model is used as a generator and a discriminator to form a generative confrontation network, and the confrontation training is performed to obtain a trained wireless channel model and a discriminator. In the application stage, use the wireless channel model to perform wireless channel-related services, use the discriminator to evaluate the output data of the wireless channel model, and determine the quality of the wireless channel model according to the evaluation result of the discriminator. When it is determined that the wireless channel model is degraded according to the evaluation result, the wireless channel model is updated. This method can use the discriminator obtained from the training settlement training to evaluate the output data of the model in the application stage, so as to perceive the reliability of the model output data in real time, and update the model in time when the model output data is unreliable.
为了更清楚地说明本申请实施例中的技术方案,下面将对实施例描述中所需要使用的附图作简单地介绍,显而易见地,下面描述中的附图仅仅是本申请的一些实施例,对于本领域普通技术人员来讲,在不付出创造性劳动的前提下,还可以根据这些附图获得其他的附图。In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings that need to be used in the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application. For those skilled in the art, other drawings can also be obtained based on these drawings without creative effort.
图1是本申请一个示例性实施例提供的神经网络模型的结构图;Fig. 1 is a structural diagram of a neural network model provided by an exemplary embodiment of the present application;
图2是本申请一个示例性实施例提供的神经网络模型的结构图;Fig. 2 is a structural diagram of a neural network model provided by an exemplary embodiment of the present application;
图3是本申请一个示例性实施例提供的无线信道模型的更新方法的流程示意图;FIG. 3 is a schematic flowchart of a method for updating a wireless channel model provided in an exemplary embodiment of the present application;
图4是本申请一个示例性实施例提供的无线信道模型的更新方法的流程示意图;FIG. 4 is a schematic flowchart of a method for updating a wireless channel model provided in an exemplary embodiment of the present application;
图5是本申请一个示例性实施例提供的无线信道模型的更新方法的流程示意图;FIG. 5 is a schematic flowchart of a method for updating a wireless channel model provided in an exemplary embodiment of the present application;
图6是本申请一个示例性实施例提供的无线信道模型的更新方法的流程示意图;FIG. 6 is a schematic flowchart of a method for updating a wireless channel model provided in an exemplary embodiment of the present application;
图7是本申请一个示例性实施例提供的无线信道模型的更新方法的实施环境示意图;FIG. 7 is a schematic diagram of an implementation environment of a method for updating a wireless channel model provided by an exemplary embodiment of the present application;
图8是本申请一个示例性实施例提供的无线信道模型的更新方法的方法流程图;FIG. 8 is a flow chart of a method for updating a wireless channel model provided in an exemplary embodiment of the present application;
图9是本申请一个示例性实施例提供的无线信道模型的更新方法的方法流程图;FIG. 9 is a flow chart of a method for updating a wireless channel model provided in an exemplary embodiment of the present application;
图10是本申请一个示例性实施例提供的无线信道模型的更新方法的流程示意图;FIG. 10 is a schematic flowchart of a method for updating a wireless channel model provided in an exemplary embodiment of the present application;
图11是本申请一个示例性实施例提供的无线信道模型的更新方法的方法流程图;FIG. 11 is a flow chart of a method for updating a wireless channel model provided in an exemplary embodiment of the present application;
图12是本申请一个示例性实施例提供的无线信道模型的更新方法的方法流程图;FIG. 12 is a flow chart of a method for updating a wireless channel model provided in an exemplary embodiment of the present application;
图13是本申请一个示例性实施例提供的更新方法的流程示意图;Fig. 13 is a schematic flowchart of an update method provided by an exemplary embodiment of the present application;
图14是本申请一个示例性实施例提供的更新方法的流程示意图;FIG. 14 is a schematic flowchart of an update method provided by an exemplary embodiment of the present application;
图15是本申请一个示例性实施例提供的无线信道模型的更新方法的方法流程图;FIG. 15 is a flow chart of a method for updating a wireless channel model provided in an exemplary embodiment of the present application;
图16是本申请一个示例性实施例提供的更新方法的流程示意图;Fig. 16 is a schematic flowchart of an update method provided by an exemplary embodiment of the present application;
图17是本申请一个示例性实施例提供的更新方法的流程示意图;Fig. 17 is a schematic flowchart of an update method provided by an exemplary embodiment of the present application;
图18是本申请一个示例性实施例提供的无线信道模型的更新方法的方法流程图;FIG. 18 is a flow chart of a method for updating a wireless channel model provided in an exemplary embodiment of the present application;
图19是本申请一个示例性实施例提供的更新方法的流程示意图;Fig. 19 is a schematic flowchart of an update method provided by an exemplary embodiment of the present application;
图20是本申请一个示例性实施例提供的更新方法的流程示意图;Fig. 20 is a schematic flowchart of an update method provided by an exemplary embodiment of the present application;
图21是本申请一个示例性实施例提供的第一通信装置的结构框图;Fig. 21 is a structural block diagram of a first communication device provided by an exemplary embodiment of the present application;
图22是本申请一个示例性实施例提供的第二通信装置的结构框图;Fig. 22 is a structural block diagram of a second communication device provided by an exemplary embodiment of the present application;
图23是本申请一个示例性实施例提供的通信设备的结构示意图。Fig. 23 is a schematic structural diagram of a communication device provided by an exemplary embodiment of the present application.
为使本申请的目的、技术方案和优点更加清楚,下面将结合附图对本申请实施方式作进一步地详细描述。In order to make the purpose, technical solution and advantages of the present application clearer, the implementation manners of the present application will be further described in detail below in conjunction with the accompanying drawings.
神经网络:人工神经网络(Artificial Neural Networks,简写为ANNs)也简称为神经网络(NNs)或称作连接模型(Connection Model),是一种模仿动物神经网络行为特征,进行分布式并行信息处理的算法数学模型。一个简单的神经网络的基本结构如图1所示,包括:输入层101,隐藏层102和输出层103。输入层101负责接收数据,隐藏层102对数据的处理,最后的结果在输出层103产生。在这其中,各个节点代表一个处理单元,也可以称为一个神经元,多个神经元组成一层神经网络,多层的信息传递与处理构造出一个整体的神经网络。Neural network: Artificial Neural Networks (ANNs for short), also referred to as neural networks (NNs) or Connection Model, is a model that imitates the behavioral characteristics of animal neural networks and performs distributed parallel information processing. Algorithmic Mathematical Model. The basic structure of a simple neural network is shown in FIG. 1 , including: an input layer 101 , a hidden layer 102 and an output layer 103 . The input layer 101 is responsible for receiving data, the hidden layer 102 processes the data, and the final result is generated in the output layer 103 . Among them, each node represents a processing unit, which can also be called a neuron. Multiple neurons form a layer of neural network, and multiple layers of information transmission and processing construct an overall neural network.
随着神经网络研究的不断发展,近年来又提出了神经网络深度学习算法,较多的隐层被引入,通过多隐层的神经网络逐层训练进行特征学习,极大地提升了神经网络的学习和处理能力,并在模式识别、信号处理、优化组合、异常探测等方面广泛被应用。With the continuous development of neural network research, neural network deep learning algorithms have been proposed in recent years, more hidden layers have been introduced, and feature learning is performed through layer-by-layer training of neural networks with multiple hidden layers, which greatly improves the learning of neural networks. And processing capabilities, and are widely used in pattern recognition, signal processing, optimization combination, anomaly detection, etc.
随着深度学习的发展,卷积神经网络(Convolutional Neural Networks,CNN)也被进一步研究。如图2所示在一个卷积神经网络中,其基本结构包括:输入层101、多个卷积层104、多个池化层105、全连接层106及输出层103。卷积层104和池化层105的引入,有效地控制了网络参数的剧增,限制了参数的个数并挖掘了局部结构的特点,提高了算法的鲁棒性。With the development of deep learning, Convolutional Neural Networks (CNN) has also been further studied. As shown in FIG. 2 , in a convolutional neural network, its basic structure includes: an input layer 101 , multiple convolutional layers 104 , multiple pooling layers 105 , a fully connected layer 106 and an output layer 103 . The introduction of the convolutional layer 104 and the pooling layer 105 effectively controls the sharp increase of network parameters, limits the number of parameters and taps the characteristics of the local structure, improving the robustness of the algorithm.
利用有限数据集训练得到的神经网络模型,在执行实际业务时,由于业务数据会随着时间推移产同程度的变化,神经网络模型将会不适用于新的业务数据,以至出现模型失效的情况。为了保障业务的正常运转,需要对神经网络模型做及时的更新。如图3所示,模型训练模块201使用数据收集模块204所收集的训练数据,训练模型,将训练好的模型部署到模型推理模块202进行应用。模型推理模块202根据数据收集模块204中的推理数据输出输出数据,将输出数据发送至业务应用模块203,用于执行业务。同时,模型推理模块202收集模型性能反馈,基于历史性能评估模型是否需要更新,向模型训练模块201反馈模型性能,使训练模块201更新模型,并向模型推理模块202发送更新后的模型。The neural network model obtained by training with a limited data set, when performing actual business, because the business data will change to the same degree over time, the neural network model will not be suitable for new business data, and even the model will fail. . In order to ensure the normal operation of the business, it is necessary to update the neural network model in time. As shown in FIG. 3 , the model training module 201 uses the training data collected by the data collection module 204 to train a model, and deploys the trained model to the model reasoning module 202 for application. The model reasoning module 202 outputs output data according to the reasoning data in the data collection module 204, and sends the output data to the business application module 203 for executing the business. At the same time, the model reasoning module 202 collects model performance feedback, evaluates whether the model needs to be updated based on historical performance, and feeds back the model performance to the model training module 201, so that the training module 201 updates the model, and sends the updated model to the model reasoning module 202.
随着神经网络的发展,神经网络在无线通信***中的应用越来越多,例如基于神经网络的CSI(Channel State Information,信道状态信息)反馈,信道估计、CSI预测等。With the development of neural networks, more and more applications of neural networks in wireless communication systems, such as neural network-based CSI (Channel State Information, channel state information) feedback, channel estimation, CSI prediction, etc.
信道估计:由于无线信道环境的复杂性和时变性,接收机针对无线信道的估计及恢复直接影响着对接收数据的恢复性能。相关技术中的信道估计及恢复过程如图4所示。发射机在时频资源上除了发射发送数据符号301外,还会发射一系列接收机已知的导频符号(发送参考信号符号302),如CSI-RS(Reference Signal,参考信号)、DMRS(Demodulation Reference Signal,解调参考信号)等。接收机接收发送数据符号301对应的接收数据符号303,以及发送后参考信号符号302对应的接收参考信号符号304,进行信道估计311:接收机根据真实导频(发送参考信号符号302)与接收导频(接收参考信号符号304)利用LS算法等方法估计出该参考信号位置上的信道信息。然后基于已估计出的信道信息进行信道恢复312:接收机根据导频位置(参考信号位置)上估计出的信道信息利用插值算法恢复出全时频资源上的信道信息,用于后续的信道信息反馈或数据恢复等。Channel estimation: due to the complexity and time-varying nature of the wireless channel environment, the receiver's estimation and recovery of the wireless channel directly affects the recovery performance of the received data. The channel estimation and restoration process in the related art is shown in FIG. 4 . In addition to transmitting the data symbol 301 on the time-frequency resource, the transmitter will also transmit a series of pilot symbols known to the receiver (reference signal symbol 302), such as CSI-RS (Reference Signal, reference signal), DMRS ( Demodulation Reference Signal, demodulation reference signal), etc. The receiver receives the received data symbol 303 corresponding to the transmitted data symbol 301, and the received reference signal symbol 304 corresponding to the transmitted reference signal symbol 302, and performs channel estimation 311: the receiver performs channel estimation 311 according to the real pilot (transmitted reference signal symbol 302) and the received pilot Frequency (received reference signal symbol 304) uses methods such as the LS algorithm to estimate the channel information at the position of the reference signal. Then perform channel recovery 312 based on the estimated channel information: the receiver uses an interpolation algorithm to recover the channel information on the full time-frequency resource according to the channel information estimated at the pilot position (reference signal position), which is used for subsequent channel information Feedback or data recovery etc.
CSI反馈:鉴于AI(Artificial Intelligence,人工智能)技术,尤其是深度学习在计算机视觉、自然语言处理等方面取得了巨大的成功,通信领域开始尝试利用深度学习来解决传统通信方法难以解决的技术难题。深度学习中常用的神经网络架构是非线性且是数据驱动的,可以对实际信道矩阵数据进行特征提取并在基站侧尽可能还原UE(User Equipment,用户设备)侧压缩反馈的信道矩阵信息,在保证还原信道信息的同时也为UE侧降低CSI反馈开销提供了可能性。基于深度学习的CSI反馈将信道信息视作待压缩图像,利用深度学习自编码器对信道信息进行压缩反馈,并在接收端对压缩后的信道图像进行重构,可以更大程度地保留信道信息。CSI feedback: In view of AI (Artificial Intelligence, artificial intelligence) technology, especially deep learning has achieved great success in computer vision, natural language processing, etc., the field of communication has begun to try to use deep learning to solve technical problems that are difficult to solve by traditional communication methods . The neural network architecture commonly used in deep learning is nonlinear and data-driven. It can extract features from the actual channel matrix data and restore the channel matrix information compressed and fed back by the UE (User Equipment) side as much as possible on the base station side. While restoring the channel information, it also provides a possibility for the UE side to reduce the CSI feedback overhead. The CSI feedback based on deep learning regards the channel information as the image to be compressed, uses the deep learning self-encoder to compress the channel information, and reconstructs the compressed channel image at the receiving end, which can preserve the channel information to a greater extent .
自编码器包括编码器和解码器,其中,编码器部署在发送端(UE侧),解码器部署在接收端(基站侧/接入网设备侧)。即,发送端通过信道估计得到信道信息后,利用编码器对信道信息进行压缩编码得到编码结果,发送端向接收端发送编码结果,即,将压缩后的比特流通过空口反馈链路反馈给接收端。接收端接收编码结果,利用解码器对编码结果进行解码得到信道信息,即,通过解码器根据反馈比特流对信道信息进行恢复,以获得完整的CSI反馈。示例性的,编码器可以采用多层全连接层的叠加,解码器可以采用卷积层与残差结构的设计。示例性的,在该编解码框架不变的情况下,编码器和解码器内部的网络模型结构可进行灵活设计。The self-encoder includes an encoder and a decoder, wherein the encoder is deployed at the sending end (UE side), and the decoder is deployed at the receiving end (base station side/access network device side). That is, after the transmitting end obtains the channel information through channel estimation, the encoder compresses and encodes the channel information to obtain the encoding result, and the transmitting end sends the encoding result to the receiving end, that is, the compressed bit stream is fed back to the receiving end through the air interface feedback link. end. The receiving end receives the encoding result and uses the decoder to decode the encoding result to obtain channel information, that is, the decoder restores the channel information according to the feedback bit stream to obtain complete CSI feedback. Exemplarily, the encoder can adopt the superposition of multiple fully connected layers, and the decoder can adopt the design of the convolutional layer and the residual structure. Exemplarily, under the condition that the encoding and decoding framework remains unchanged, the internal network model structure of the encoder and decoder can be flexibly designed.
CSI预测:将收端接收的周期性CSI反馈信息输入神经网络模型,能够预测非测量时刻的CSI,从而获得更完整的CSI。如图5所示,在阴影方框305对应的时刻,接收端接收周期性的CSI反馈,而在白色方框306以及黑色方框307对应的时刻,则没有CSI反馈,其中黑色方框307代表未来数个时刻。将阴影方框305对应的时刻接收到的CSI反馈输入神经网络模型,从而输出预测得到的黑色方框307对应的时刻的CSI。CSI prediction: Input the periodic CSI feedback information received by the receiving end into the neural network model, which can predict the CSI at non-measurement time, so as to obtain a more complete CSI. As shown in Figure 5, at the moment corresponding to the shaded box 305, the receiving end receives periodic CSI feedback, and at the moment corresponding to the white box 306 and black box 307, there is no CSI feedback, where the black box 307 represents several moments in the future. The CSI received at the time corresponding to the shaded box 305 is fed back into the neural network model, so as to output the predicted CSI at the time corresponding to the black box 307 .
生成对抗网络(Generative Adversarial Networks,GAN)是通过对抗训练的方式来使得生成器产生的样本服从真实数据分布。Generative Adversarial Networks (GAN) is to make the samples generated by the generator obey the real data distribution through confrontation training.
在生成对抗网络中,有两个网络进行对抗训练。一个是判别网络(判别器),判别器的目标是尽量准确地区分输入数据是真实数据还是由生成器产生的数据;另一个是生成网络(生成器),生成器的目标是尽量生成判别器无法区分的样本。这两个目标相反的网络不断地进行交替训练。当模型收敛时,如果判别器再也无法判断生成器输出的数据是否为真实数据,则生成器可以生成符合真实数据分布的样本。In Generative Adversarial Networks, two networks are trained against each other. One is the discrimination network (discriminator), the goal of the discriminator is to distinguish as accurately as possible whether the input data is real data or the data generated by the generator; the other is the generation network (generator), the goal of the generator is to generate the discriminator as much as possible Indistinguishable samples. These two networks with opposite goals are constantly alternately trained. When the model converges, if the discriminator can no longer judge whether the data output by the generator is real data, the generator can generate samples that conform to the real data distribution.
生成对抗网络是一种神经网络模型的训练方法,基于图6对生成对抗网络的一种对抗训练方法进行简单说明。生成对抗网络由生成器308和判别器309组成。Generative confrontation network is a training method of neural network model. Based on Figure 6, a confrontation training method of generative confrontation network is briefly described. The Generative Adversarial Network consists of a generator 308 and a discriminator 309 .
在训练阶段,首先固定生成器中的模型参数,训练判别器:将输入数据输入生成器得到生成数据。将生成器输出的生成数据,以及样本数据集中的真实数据,分别作为判别器的输入,分别输入判别器得到预测标签,基于预测标签与实际标签计算损失值。其中,当输入判别器的数据是生成数据时,实际标签为生 成标签,当输入判别器的数据是真实数据时,实际标签为真实标签。基于预测标签与实际标签的误差(损失值)来调整判别器中的模型参数、训练判别器,使判别器输出的预测标签贴近实际标签,即,使判别器能够准确判别出输入的数据是真实数据还是生成数据。In the training phase, the model parameters in the generator are first fixed, and the discriminator is trained: the input data is input into the generator to obtain generated data. The generated data output by the generator and the real data in the sample data set are respectively used as the input of the discriminator, which are respectively input into the discriminator to obtain the predicted label, and the loss value is calculated based on the predicted label and the actual label. Among them, when the data input to the discriminator is generated data, the actual label is the generated label, and when the data input to the discriminator is real data, the actual label is the real label. Adjust the model parameters in the discriminator and train the discriminator based on the error (loss value) between the predicted label and the actual label, so that the predicted label output by the discriminator is close to the actual label, that is, the discriminator can accurately judge that the input data is real Data is still data.
在判别器收敛后,固定判别器中的模型参数,训练生成器。在训练生成器时,将生成器和判别器作为一个整体,将输入数据输入生成器得到生成数据,将生成数据输入判别器得到预测标签,然后基于预测标签与真实标签计算损失值,基于预测标签与真实标签的误差(损失值)来调整生成器中的模型参数、训练生成器,即,使生成器输出的数据能够被判别器判别为真实数据,则表示生成器生成的生成数据贴近真实数据。After the discriminator converges, the model parameters in the discriminator are fixed and the generator is trained. When training the generator, the generator and the discriminator are taken as a whole, the input data is input into the generator to obtain the generated data, the generated data is input into the discriminator to obtain the predicted label, and then the loss value is calculated based on the predicted label and the real label, based on the predicted label The error (loss value) with the real label is used to adjust the model parameters in the generator and train the generator, that is, to make the data output by the generator be recognized as real data by the discriminator, which means that the generated data generated by the generator is close to the real data .
示例性的,可以按照上述方法,不断地固定一个网络的模型参数,训练另一个网络,迭代训练使生成器收敛,最终得到训练好的生成器。训练好的生成器可以基于输入数据输出贴近真实数据的生成数据,而判别器则充分学习了真实数据的特征。Exemplarily, according to the above method, the model parameters of one network can be continuously fixed, another network can be trained, and the generator can be converged through iterative training, and finally a trained generator can be obtained. The trained generator can output generated data close to the real data based on the input data, while the discriminator fully learns the characteristics of the real data.
相关技术中,神经网络无线信道模型的更新模式大多基于推理侧的模型性能反馈,将性能指标量化后判断是否需要更新当前的神经网络模型。即,相关技术中的模型更新是通过反馈数据来驱动的。In related technologies, the update mode of the wireless channel model of the neural network is mostly based on the model performance feedback at the inference side, and the performance index is quantified to determine whether the current neural network model needs to be updated. That is, model updating in the related art is driven by feedback data.
采用相关技术中的方法实时反馈模型性能数据,其空口传输的资源开销较大。而周期性反馈性能数据可以减小模型性能反馈带来的开销,但是牺牲了对模型劣化的实时感知能力。The method in the related art is used to feed back the model performance data in real time, and the resource overhead of the air interface transmission is relatively large. The periodic feedback of performance data can reduce the overhead caused by model performance feedback, but sacrifices the real-time perception of model degradation.
此外,模型性能的反馈需要经历多个模块,从模型输出发送到具体业务模块指导业务决策,同时业务模块承担量化模型可靠性的角色,通过业务指标间接反馈模型性能。这种方法进行模型性能反馈要涉及多网元交互,反馈流程繁琐,而且仅仅是间接地反映神经网络模型的性能,属于比较粗粒度的反馈。In addition, the feedback of model performance needs to go through multiple modules, and the output of the model is sent to specific business modules to guide business decisions. At the same time, the business module assumes the role of quantifying the reliability of the model and indirectly feeds back the model performance through business indicators. This method of model performance feedback involves the interaction of multiple network elements, the feedback process is cumbersome, and it only indirectly reflects the performance of the neural network model, which belongs to relatively coarse-grained feedback.
并且,上述模型性能反馈的方案,需要连续观测一段时间的性能反馈,来粗略量化神经网络模型的劣化程度,定性判断模型是否失效。这样的反馈方案并不会指导模型如何进行更新。反馈形式单一,不具有多元化。面对复杂的场景,模型可能无法及时更新,更无法根据数据分布的变化而进行动态的调整,这使得模型陷入更新赶不上数据变化的困局,影响模型的长期持续性能表现。Moreover, the above-mentioned model performance feedback scheme requires continuous observation of performance feedback for a period of time to roughly quantify the degree of degradation of the neural network model and qualitatively determine whether the model is invalid. Such a feedback scheme does not guide how the model should be updated. The form of feedback is single and not diversified. In the face of complex scenarios, the model may not be updated in time, let alone dynamically adjusted according to changes in data distribution, which makes the model fall into a dilemma where the update cannot keep up with the data changes, affecting the long-term continuous performance of the model.
进一步地,神经网络模型大多是单一神经网络模型或者多个同样功能的神经网络模型融合,神经网络模型只负责预测输出,并不输出可靠性指示,无法实时评估当前输出结果是否可靠,从而导致基于神经网络模型的输出数据指导业务模块决策带有一定的盲目性,上述的模型性能反馈方案也属于事后反馈,属于补救性措施。即,上述的模型更新方案是牺牲业务性能前提下的补救性更新方案。Furthermore, most of the neural network models are a single neural network model or the fusion of multiple neural network models with the same function. The neural network model is only responsible for predicting the output, and does not output reliability indicators. It is impossible to evaluate the reliability of the current output results in real time, resulting in The output data of the neural network model guides the decision-making of the business module with a certain degree of blindness. The above-mentioned model performance feedback scheme also belongs to post-event feedback and is a remedial measure. That is, the above-mentioned model update solution is a remedial update solution under the premise of sacrificing business performance.
基于深度学习的模型性能与数据分布强相关,由于无线环境不稳定,数据分布难免也会受时间、环境、***策略等因素的影响。因此,神经网络模型无法避免失效的情况。The performance of the model based on deep learning is strongly related to the data distribution. Due to the unstable wireless environment, the data distribution will inevitably be affected by factors such as time, environment, and system strategy. Therefore, neural network models cannot avoid failures.
综上所述,如何有效、快速地更新神经网络模型来应对各种不可控因素带来的数据分布变化,以及如何将事后补救性模型更新方案向及时感知模型劣化、及时主动更新方案演进的研究,是亟待解决的模型泛化问题。To sum up, how to effectively and quickly update the neural network model to deal with the changes in data distribution caused by various uncontrollable factors, and how to evolve the post-event remedial model update scheme to the timely perception of model degradation and timely active update scheme evolution , is an urgent model generalization problem to be solved.
针对上述问题,本申请实施例提供了一种无线信道模型的更新方法,能够及时感知模型劣化,及时主动进行模型更新。In view of the above problems, an embodiment of the present application provides a method for updating a wireless channel model, which can sense model degradation in a timely manner and actively update the model in a timely manner.
本申请实施例描述的网络架构以及业务场景是为了更加清楚地说明本申请实施例的技术方案,并不构成对本申请实施例提供的技术方案的限定,本领域普通技术人员可知,随着网络架构的演变和新业务场景的出现,本申请实施例提供的技术方案对于类似的技术问题,同样适用。The network architecture and business scenarios described in the embodiments of the present application are for more clearly illustrating the technical solutions of the embodiments of the present application, and do not constitute limitations on the technical solutions provided by the embodiments of the present application. The evolution of the technology and the emergence of new business scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.
请参考图7,其示出了本申请一个实施例提供的网络架构100的示意图。该网络架构100可以包括:终端设备10和网络设备,其中,网络设备可以包括接入网设备20和核心网设备30中的至少一个。Please refer to FIG. 7 , which shows a schematic diagram of a network architecture 100 provided by an embodiment of the present application. The network architecture 100 may include: a terminal device 10 and a network device, where the network device may include at least one of an access network device 20 and a core network device 30 .
终端设备10可以指用户设备(User Equipment,UE)、接入终端、用户单元、用户站、移动站、移动台、远方站、远程终端、移动设备、无线通信设备、用户代理或用户装置。可选地,终端设备10还可以是蜂窝电话、无绳电话、会话启动协议(Session Initiation Protocol,SIP)电话、无线本地环路(Wireless Local Loop,WLL)站、个人数字处理(Personal Digita1 Assistant,PDA)、具有无线通信功能的手持设备、计算设备或连接到无线调制解调器的其它处理设备、车载设备、可穿戴设备,第五代移动通信***(5th Generation System,5GS)中的终端设备或者未来演进的公用陆地移动通信网络(Pub1ic Land Mobi1e Network,PLMN)中的终端设备等,本申请实施例对此并不限定。为方便描述,上面提到的设备统称为终端设备。终端设备10的数量通常为多个,每一个接入网设备20所管理的小区内可以分布一个或多个终端设备10。The terminal equipment 10 may refer to a user equipment (User Equipment, UE), an access terminal, a subscriber unit, a subscriber station, a mobile station, a mobile station, a remote station, a remote terminal, a mobile device, a wireless communication device, a user agent or a user device. Optionally, the terminal device 10 may also be a cellular phone, a cordless phone, a Session Initiation Protocol (Session Initiation Protocol, SIP) phone, a Wireless Local Loop (Wireless Local Loop, WLL) station, a Personal Digital Assistant (PDA) ), handheld devices with wireless communication functions, computing devices or other processing devices connected to wireless modems, vehicle-mounted devices, wearable devices, terminal devices in the fifth generation mobile communication system (5th Generation System, 5GS) or future evolution The terminal equipment in the Public Land Mobile Network (Public Land Mobile Network, PLMN), etc., is not limited in this embodiment of the present application. For convenience of description, the devices mentioned above are collectively referred to as terminal devices. The number of terminal devices 10 is generally multiple, and one or more terminal devices 10 may be distributed in a cell managed by each access network device 20 .
接入网设备20是一种部署在接入网中用以为终端设备10提供无线通信功能的设备。接入网设备20可以包括各种形式的宏基站,微基站,中继站,接入点等等。在采用不同的无线接入技术的***中,具备接入网设备功能的设备的名称可能会有所不同,例如在5G NR***中,称为gNodeB或者gNB。随着通信技术的演进,“接入网设备”这一名称可能会变化。为方便描述,本申请实施例中,上述为终端设备10提供无线通信功能的装置统称为接入网设备。可选地,通过接入网设备20,终端设备10和核心网设备30之间 可以建立通信关系。示例性地,在长期演进(Long Term Evolution,LTE)***中,接入网设备20可以是演进的通用陆地无线网(Evolved Universal Terrestrial Radio Access Network,EUTRAN)或者EUTRAN中的一个或者多个eNodeB;在5G NR***中,接入网设备20可以是无线接入网(Radio Access Network,RAN)或者RAN中的一个或者多个gNB。在本申请实施例中,所述的网络设备除特别说明之外,是指接入网设备20,如基站。The access network device 20 is a device deployed in an access network to provide a wireless communication function for the terminal device 10 . The access network device 20 may include various forms of macro base stations, micro base stations, relay stations, access points, and so on. In systems using different wireless access technologies, the names of devices with access network device functions may be different. For example, in 5G NR systems, they are called gNodeB or gNB. With the evolution of communication technology, the name "access network equipment" may change. For the convenience of description, in the embodiment of the present application, the above-mentioned devices that provide the wireless communication function for the terminal device 10 are collectively referred to as access network devices. Optionally, through the access network device 20, a communication relationship can be established between the terminal device 10 and the core network device 30. Exemplarily, in a Long Term Evolution (LTE) system, the access network device 20 may be an Evolved Universal Terrestrial Radio Access Network (Evolved Universal Terrestrial Radio Access Network, EUTRAN) or one or more eNodeBs in EUTRAN; In the 5G NR system, the access network device 20 may be a radio access network (Radio Access Network, RAN) or one or more gNBs in the RAN. In the embodiment of the present application, unless otherwise specified, the network device refers to the access network device 20, such as a base station.
核心网设备30是部署在核心网中的设备,核心网设备30的功能主要是提供用户连接、对用户的管理以及对业务完成承载,作为承载网络提供到外部网络的接口。例如,5G NR***中的核心网设备可以包括接入和移动性管理功能(Access and Mobility Management Function,AMF)网元、用户平面功能(User Plane Function,UPF)网元和会话管理功能(Session Management Function,SMF)网元等。The core network device 30 is a device deployed in the core network. The functions of the core network device 30 are mainly to provide user connections, manage users, and carry out services, and provide an interface to external networks as a bearer network. For example, the core network equipment in the 5G NR system can include access and mobility management function (Access and Mobility Management Function, AMF) network element, user plane function (User Plane Function, UPF) network element and session management function (Session Management Function) Function, SMF) network element, etc.
在一个示例中,接入网设备20与核心网设备30之间通过某种空口技术互相通信,例如5G NR***中的NG接口。接入网设备20与终端设备10之间通过某种空口技术互相通信,例如Uu接口。In an example, the access network device 20 and the core network device 30 communicate with each other through a certain air interface technology, such as the NG interface in the 5G NR system. The access network device 20 and the terminal device 10 communicate with each other through a certain air interface technology, such as a Uu interface.
本公开实施例中的“5G NR***”也可以称为5G***或者NR***,但本领域技术人员可以理解其含义。本公开实施例描述的技术方案可以适用于5G NR***,也可以适用于5G NR***后续的演进***。The "5G NR system" in the embodiments of the present disclosure may also be called a 5G system or an NR system, but those skilled in the art can understand its meaning. The technical solution described in the embodiments of the present disclosure can be applied to the 5G NR system, and can also be applied to the subsequent evolution system of the 5G NR system.
图8示出了本申请一个示例性实施例提供的无线信道模型的更新方法的流程图。本实施例以该方法应用于第一通信设备中来举例说明,例如,第一通信设备可以是图1所示的终端10或网络设备20。该方法包括:Fig. 8 shows a flowchart of a method for updating a wireless channel model provided by an exemplary embodiment of the present application. This embodiment is described by taking the method applied to a first communication device as an example. For example, the first communication device may be the terminal 10 or the network device 20 shown in FIG. 1 . The method includes:
步骤410:通过第一无线信道模型得到第一信道信息,第一无线信道模型为第一生成对抗网络中的生成器。Step 410: Obtain first channel information through a first wireless channel model, where the first wireless channel model is a generator in the first generative adversarial network.
第一无线信道模型是作为生成器与第一判别器对抗训练得到的神经网络模型。The first wireless channel model is a neural network model that is trained as a generator against a first discriminator.
本申请实施例以该无线信道模型的更新方法应用于无线信道模型为例进行说明,该方法还可以应用于通信领域中其他神经网络模型,或,其他领域(例如,图像处理、自然语言处理)中的神经网络模型。当应用于其他神经网络模型时,可以使用其他神经网络模型的名称替换本申请实施例中的“无线信道模型”。In the embodiment of the present application, the method for updating the wireless channel model is applied to the wireless channel model as an example for illustration. This method can also be applied to other neural network models in the communication field, or other fields (for example, image processing, natural language processing) The neural network model in . When applied to other neural network models, the names of other neural network models can be used to replace the "wireless channel model" in the embodiment of the present application.
示例性的,步骤410为应用训练好的第一无线信道模型的步骤,在模型应用过程中使用第一无线信道模型基于第二信道信息得到第一信道信息。使用第一信道信息执行无线信道相关的业务,或指导无线信道相关业务的决策。Exemplarily, step 410 is a step of applying the trained first wireless channel model. During the model application process, the first wireless channel model is used to obtain the first channel information based on the second channel information. Use the first channel information to perform wireless channel-related services, or guide decisions on wireless channel-related services.
第一无线信道模型、第二无线信道模型仅用于区分更新前的无线信道模型和更新后的无线信道模型,二者都可以称为无线信道模型。The first radio channel model and the second radio channel model are only used to distinguish the radio channel model before updating and the radio channel model after updating, and both may be referred to as radio channel models.
无线信道模型的网络结构可以是任意的,本申请实施例对此不加以限定。The network structure of the wireless channel model may be arbitrary, which is not limited in this embodiment of the present application.
示例性的,通过第一无线信道模型基于第二信道信息得到第一信道信息,是指:第二信道信息是输入数据,第一信道信息是输出数据,将第二信道信息输入第一无线信道模型,第一无线信道模型基于输入的第二信道信息输出第一信道信息。Exemplarily, obtaining the first channel information based on the second channel information through the first wireless channel model refers to: the second channel information is input data, the first channel information is output data, and the second channel information is input into the first wireless channel The first wireless channel model outputs the first channel information based on the input second channel information.
第二信道信息是输入无线信道模型的信道信息。第一信道信息是无线信道模型基于第二信道信息输出的信道信息。示例性的,第二信道信息和第一信道信息为同类型的信道信息,或,不同类型的信道信息。The second channel information is channel information input into the wireless channel model. The first channel information is channel information output by the wireless channel model based on the second channel information. Exemplarily, the second channel information and the first channel information are the same type of channel information, or different types of channel information.
信道信息(第二信道信息或第一信道信息)为信道相关的数据,例如,信道信息可以包括如下参数中的至少一种:发射天线、接收天线、时延、FDM(Frequency Division Multiplexing,频分复用)符号数、子载波数量、信道特征值分解后得到的特征向量等。Channel information (second channel information or first channel information) is channel-related data, for example, channel information may include at least one of the following parameters: transmitting antenna, receiving antenna, time delay, FDM (Frequency Division Multiplexing, frequency division Multiplexing) number of symbols, number of subcarriers, eigenvectors obtained after channel eigenvalue decomposition, etc.
在训练阶段,第一无线信道模型与第一判别器组成第一生成对抗网络,其中,第一无线信道模型作为第一生成对抗网络中的生成器,采用生成对抗网络的对抗训练方法训练第一无线信道模型和第一判别器。In the training phase, the first wireless channel model and the first discriminator form the first generative adversarial network, wherein the first wireless channel model is used as the generator in the first generative adversarial network, and the first Wireless channel model and first discriminator.
步骤420:在通过第一判别器基于第一信道信息确定第一无线信道模型满足更新条件的情况下,触发更新第一无线信道模型,第一判别器为第一生成对抗网络中的判别器。Step 420: When the first discriminator determines that the first wireless channel model satisfies the update condition based on the first channel information, trigger an update of the first wireless channel model. The first discriminator is a discriminator in the first GAN.
示例性的,使用第一判别器来确定第一无线信道模型是否劣化。使用第一判别器来评价第一无线信道模型输出的第一信道信息,得到第一概率值,若第一概率值满足更新条件,则表示第一无线信道模型满足更新条件,触发更新第一无线信道模型。Exemplarily, a first discriminator is used to determine whether the first wireless channel model is degraded. Use the first discriminator to evaluate the first channel information output by the first wireless channel model to obtain the first probability value. If the first probability value satisfies the update condition, it means that the first wireless channel model meets the update condition, and triggers to update the first wireless channel model. channel model.
通过第一判别器评价第一信道信息得到第一概率值(第一评价结果)。The first probability value (first evaluation result) is obtained by evaluating the first channel information by the first discriminator.
将第一信道信息输入第一判别器,得到第一概率值。示例性的,第一概率值为0-1的数值,或,第一概率值为0或1。Input the first channel information into the first discriminator to obtain the first probability value. Exemplarily, the first probability value is a value of 0-1, or, the first probability value is 0 or 1.
在训练阶段,第一判别器被训练为用于判别输入的数据是否为真实数据。利用第一判别器来对抗训练生成器(第一无线信道模型)使生成器能够输出贴近真实信道信息的信道信息。In the training phase, the first discriminator is trained to determine whether the input data is real data. Utilizing the first discriminator to adversarially train the generator (the first wireless channel model) enables the generator to output channel information close to real channel information.
在步骤420中的模型应用阶段,继续应用第一判别器来判别第一信道信息的可靠性。第一概率值越接近真实数据则说明第一信道信息越可靠,反之第一概率值越接近非真实数据则说明第一信道信息越不可靠。In the model application stage in step 420, continue to apply the first discriminator to judge the reliability of the first channel information. The closer the first probability value is to real data, the more reliable the first channel information is, whereas the closer the first probability value is to non-real data, the less reliable the first channel information is.
需要注意的是,在相关技术中,判别器仅在训练阶段用于与生成器的对抗训练,在模型的应用阶段则 丢弃判别器仅使用生成器执行业务。而在本申请实施例提供的方法中,创造性地将判别器在应用阶段用于评价生成器的输出结果,从而能够基于判别器输出的评价结果实时观测生成器是否劣化。It should be noted that in related technologies, the discriminator is only used for confrontation training with the generator in the training phase, and in the application phase of the model, the discriminator is discarded and only the generator is used to perform business. However, in the method provided by the embodiment of the present application, the discriminator is creatively used to evaluate the output result of the generator in the application stage, so that whether the generator is degraded can be observed in real time based on the evaluation result output by the discriminator.
第一判别器、第二判别器仅用于区分更新前的判别器和更新后的判别器,二者都可以称为判别器。其中,第一判别器是与第一无线信道模型对抗训练得到的判别器,第二判别器是用于与第二无线信道模型对抗训练的判别器,或,第二判别器是与第二无线信道模型对抗训练得到的判别器。The first discriminator and the second discriminator are only used to distinguish the pre-update discriminator from the updated discriminator, both of which can be called discriminators. Wherein, the first discriminator is a discriminator trained against the first wireless channel model, and the second discriminator is a discriminator used for confrontation training with the second wireless channel model, or the second discriminator is trained against the second wireless channel model. The channel model is trained against the discriminator.
示例性的,判别器是一种二分类模型。例如,判别器的输出为0或1,其中,0代表一类,1代表另一类。例如,0代表非真实数据,1代表真实数据。或者,判别器的输出为0到1的数值,其中,数值越接近0则表示输入判别器的数据越接近非真实数据,数值越接近1则表示输入判别器的数据越接近真实数据,0.5则表示无法判断输入判别器的数据是真实数据还是非真实数据。Exemplarily, the discriminator is a binary classification model. For example, the output of the discriminator is 0 or 1, where 0 represents one class and 1 represents the other. For example, 0 represents non-real data and 1 represents real data. Alternatively, the output of the discriminator is a value from 0 to 1, where the closer the value is to 0, the closer the data input to the discriminator is to the unreal data, the closer the value is to 1, the closer the data input to the discriminator is to real data, and 0.5 means Indicates that it is impossible to judge whether the data input to the discriminator is real data or non-real data.
在基于第一概率值确定第一无线信道模型满足更新条件的情况下,触发更新第一无线信道模型。In a case where it is determined based on the first probability value that the first wireless channel model satisfies the update condition, an update of the first wireless channel model is triggered.
根据第一概率值确定第一无线信道模型的劣化程度,在根据第一概率值确定第一无线信道模型满足更新条件的情况下,即,第一无线信道模型的性能较差时,则更新第一无线信道模型得到第二无线信道模型。Determine the degree of degradation of the first wireless channel model according to the first probability value, and if it is determined according to the first probability value that the first wireless channel model satisfies the update condition, that is, when the performance of the first wireless channel model is poor, update the first wireless channel model A radio channel model obtains a second radio channel model.
示例性的,更新条件用于判定第一无线信道模型是否劣化,当第一无线信道模型输出的第一信道信息,较多地被第一判别器判别为非真实数据(即,概率值为0或概率值低于0.5)时,则说明第一无线信道模型已经劣化,无法输出较为真实的信道信息,则第一无线信道模型需要进行更新。也即,劣化是指第一无线信道模型输出的信道信息失去真实性/可靠性。Exemplarily, the update condition is used to determine whether the first wireless channel model is degraded. When the first channel information output by the first wireless channel model is mostly judged as unreal data by the first discriminator (that is, the probability value is 0 or the probability value is lower than 0.5), it means that the first wireless channel model has deteriorated and cannot output relatively real channel information, and the first wireless channel model needs to be updated. That is, degradation means that the channel information output by the first wireless channel model loses authenticity/reliability.
由于输入第一无线信道模型的信道信息(第二信道信息)的数据布局变化,第一无线信道模型可能无法准确提取新的第二信道信息中的特征,无法输出较为真实的信道信息,此时需要更新第一无线信道模型,使之重新学习数据布局变化后的第二信道信息中的特征,输出较为真实的信道数据。Due to the change in the data layout of the channel information (second channel information) input to the first wireless channel model, the first wireless channel model may not be able to accurately extract the features in the new second channel information, and cannot output more realistic channel information. At this time It is necessary to update the first wireless channel model so that it can relearn the features in the second channel information after the data layout changes, and output more realistic channel data.
更新第一无线信道模型是指继续训练第一无线信道模型得到第二无线信道模型。即,使用新的训练样本训练第一无线信道模型得到第二无线信道模型。示例性的,训练样本中包括第二信道信息,即,用应用过程中输入第一无线信道模型的真实信道信息来训练第一无线信道模型,使之学习第二信道信息的特征,适应数据变化。Updating the first wireless channel model refers to continuing to train the first wireless channel model to obtain the second wireless channel model. That is, new training samples are used to train the first wireless channel model to obtain the second wireless channel model. Exemplarily, the training samples include the second channel information, that is, use the real channel information input to the first wireless channel model during the application process to train the first wireless channel model, so that it can learn the characteristics of the second channel information and adapt to data changes .
综上所述,本实施例提供的方法,在训练无线信道模型时,将无线信道模型作为生成器与判别器组成生成对抗网络,进行对抗训练,得到训练好的无线信道模型和判别器。在应用阶段,使用无线信道模型执行无线信道相关的业务,并使用判别器评价无线信道模型输出的数据,根据判别器的评价结果来确定无线信道模型的优劣。当根据评价结果确定无线信道模型劣化时,对无线信道模型进行更新。该方法可以利用训练结算训练得到的判别器,在应用阶段对模型的输出数据进行评价,从而实时感知模型输出数据的可靠性,在模型输出数据不可靠时及时对模型进行更新。To sum up, in the method provided in this embodiment, when training the wireless channel model, the wireless channel model is used as a generator and a discriminator to form an adversarial network, and the adversarial training is performed to obtain a trained wireless channel model and a discriminator. In the application stage, use the wireless channel model to perform wireless channel-related services, use the discriminator to evaluate the output data of the wireless channel model, and determine the quality of the wireless channel model according to the evaluation result of the discriminator. When it is determined that the wireless channel model is degraded according to the evaluation result, the wireless channel model is updated. This method can use the discriminator obtained from the training settlement training to evaluate the output data of the model in the application stage, so as to perceive the reliability of the model output data in real time, and update the model in time when the model output data is unreliable.
图9示出了本申请一个示例性实施例提供的无线信道模型的更新方法的流程图。本实施例以该方法应用于第二通信设备中来举例说明,例如,第二通信设备可以是图1所示的终端10或网络设备20。该方法包括:Fig. 9 shows a flowchart of a method for updating a wireless channel model provided by an exemplary embodiment of the present application. This embodiment is described by taking the method applied to a second communication device as an example. For example, the second communication device may be the terminal 10 or the network device 20 shown in FIG. 1 . The method includes:
步骤310:在通过第一判别器基于第一信道信息确定第一无线信道模型满足更新条件的情况下,更新第一无线信道模型,第一信道信息是第一无线信道模型输出的,第一无线信道模型为第一生成对抗网络中的生成器,第一判别器为第一生成对抗网络中的判别器。Step 310: When the first discriminator determines that the first wireless channel model satisfies the update condition based on the first channel information, update the first wireless channel model, the first channel information is output by the first wireless channel model, and the first wireless channel model The channel model is a generator in the first generation adversarial network, and the first discriminator is a discriminator in the first generation adversarial network.
示例性的,第二通信设备与第一通信设备合作执行无线信道业务。Exemplarily, the second communication device cooperates with the first communication device to perform the wireless channel service.
1、在一种情况下,第二通信设备向第一通信设备发送第二信道信息,第一通信设备将第二信道信息输入第一无线信道模型得到第一信道信息,将第一信道信息应用于无线信道业务决策。第二信道信息是第一无线信道模型生成第一信道信息时的输入信息。1. In one case, the second communication device sends the second channel information to the first communication device, and the first communication device inputs the second channel information into the first wireless channel model to obtain the first channel information, and applies the first channel information to for wireless channel business decisions. The second channel information is input information when the first wireless channel model generates the first channel information.
此时,可以由第一通信设备使用第一判别器判别第一无线信道模型是否需要更新,并由第一通信设备进行第一无线信道模型的更新。At this time, the first communication device may use the first discriminator to determine whether the first radio channel model needs to be updated, and the first communication device may update the first radio channel model.
也可以由第一通信设备使用第一判别器判别第一无线信道模型是否需要更新,并在需要更新时,向第二通信设备发送更新指示,指示第二通信设备更新第一无线信道模型得到第二无线信道模型,并由第二通信设备向第一通信设备发送第二无线信道模型。It is also possible for the first communication device to use the first discriminator to determine whether the first wireless channel model needs to be updated, and when it needs to be updated, send an update instruction to the second communication device, instructing the second communication device to update the first wireless channel model to obtain the first wireless channel model. Two wireless channel models, and the second wireless channel model is sent by the second communication device to the first communication device.
2、在另一种情况下,第一无线信道模型包括编码器和解码器两部分。编码器和解码器分别部署在第一通信设备和第二通信设备中。2. In another case, the first wireless channel model includes two parts: an encoder and a decoder. The encoder and the decoder are respectively deployed in the first communication device and the second communication device.
此时,可以由编码器侧使用第一判别器判别第一无线信道模型是否需要更新,并由编码器侧更新第一无线信道模型,将更新后得到的第二解码器和第二判别器发送给解码器侧。At this time, the encoder side can use the first discriminator to judge whether the first wireless channel model needs to be updated, and the encoder side can update the first wireless channel model, and send the updated second decoder and second discriminator to the decoder side.
也可以由解码器侧使用第一判别器判别第一无线信道模型是否需要更新,并由解码器侧向编码器侧发送更新指示,指示编码器侧更新第一无线信道模型,并将更新后得到的第二解码器和第二判别器发送给解码器侧。It is also possible for the decoder side to use the first discriminator to judge whether the first wireless channel model needs to be updated, and the decoder side to send an update instruction to the encoder side, instructing the encoder side to update the first wireless channel model, and obtain after the update The second decoder and the second discriminator are sent to the decoder side.
也可以由编码器侧使用第一判别器判别第一无线信道模型是否需要更新,并由编码器侧向解码器侧发 送更新指示,指示解码器侧更新第一无线信道模型,并将更新后得到的第二编码器、第二解码器和第二判别器发送给编码器侧。It is also possible for the encoder side to use the first discriminator to determine whether the first wireless channel model needs to be updated, and the encoder side to send an update instruction to the decoder side, instructing the decoder side to update the first wireless channel model, and obtain after the update The second encoder, the second decoder and the second discriminator are sent to the encoder side.
也可以由解码器侧使用第一判别器判别第一无线信道模型是否需要更新,并由解码器侧更新第一无线信道模型,并将更新后得到的第二编码器发送给编码器侧。It is also possible that the decoder side uses the first discriminator to judge whether the first radio channel model needs to be updated, and the decoder side updates the first radio channel model, and sends the updated second encoder to the encoder side.
示例性的,针对第一无线信道模型只部署在单侧通信设备时,给出四种第一通信设备和第二通信设备的分工情况的列举。Exemplarily, when the first wireless channel model is only deployed on a single-side communication device, four cases of division of labor between the first communication device and the second communication device are enumerated.
情况1:第一无线信道模型部署在第一通信设备侧,第一通信设备判别模型劣化,并执行无线信道模型的更新。Case 1: The first wireless channel model is deployed on the side of the first communication device, and the first communication device determines that the model is degraded, and performs an update of the wireless channel model.
第二通信设备向第一通信设备发送第二信道信息,第一通信设备接收第二通信设备发送的第二信道信息。第一通信设备通过第一无线信道模型基于第二信道信息得到第一信道信息。第一通信设备在通过第一判别器基于第一信道信息确定第一无线信道模型满足更新条件的情况下,更新第一无线信道模型。The second communication device sends the second channel information to the first communication device, and the first communication device receives the second channel information sent by the second communication device. The first communication device obtains the first channel information based on the second channel information through the first wireless channel model. The first communication device updates the first wireless channel model when it is determined by the first discriminator based on the first channel information that the first wireless channel model satisfies an update condition.
情况2:第一无线信道模型部署在第一通信设备侧,第一通信设备判别模型劣化,并指示第二通信设备执行无线信道模型的更新。Case 2: The first wireless channel model is deployed on the side of the first communication device, and the first communication device determines that the model is degraded, and instructs the second communication device to update the wireless channel model.
第二通信设备向第一通信设备发送第二信道信息,第一通信设备接收第二通信设备发送的第二信道信息。第一通信设备通过第一无线信道模型基于第二信道信息得到第一信道信息。第一通信设备在通过第一判别器基于第一信道信息确定第一无线信道模型满足更新条件的情况下,向第二通信设备发送更新指示。第二通信设备接收更新指示,更新第一无线信道模型得到第二无线信道模型,第二通信设备向第一通信设备发送第二无线信道模型和第二判别器。第一通信设备接收并部署第二无线信道模型和第二判别器。The second communication device sends the second channel information to the first communication device, and the first communication device receives the second channel information sent by the second communication device. The first communication device obtains the first channel information based on the second channel information through the first wireless channel model. The first communication device sends an update instruction to the second communication device when it is determined by the first discriminator based on the first channel information that the first wireless channel model satisfies the update condition. The second communication device receives the update instruction, updates the first wireless channel model to obtain the second wireless channel model, and sends the second wireless channel model and the second discriminator to the first communication device. The first communications device receives and deploys a second wireless channel model and a second discriminator.
情况3:第一无线信道模型部署在第一通信设备侧,第二通信设备判别模型劣化,并执行无线信道模型的更新。Case 3: the first wireless channel model is deployed on the side of the first communication device, and the second communication device determines that the model is degraded, and performs an update of the wireless channel model.
第二通信设备向第一通信设备发送第二信道信息,第一通信设备接收第二通信设备发送的第二信道信息。第一通信设备通过第一无线信道模型基于第二信道信息得到第一信道信息。第二通信设备也通过第一无线信道模型基于第二信道信息得到第一信道信息。第二通信设备在通过第一判别器基于第一信道信息确定第一无线信道模型满足更新条件的情况下,更新第一无线信道模型。第二通信设备向第一通信设备发送第二无线信道模型和第二判别器。第一通信设备接收并部署第二无线信道模型和第二判别器。The second communication device sends the second channel information to the first communication device, and the first communication device receives the second channel information sent by the second communication device. The first communication device obtains the first channel information based on the second channel information through the first wireless channel model. The second communication device also obtains the first channel information based on the second channel information through the first wireless channel model. The second communication device updates the first wireless channel model when it is determined by the first discriminator based on the first channel information that the first wireless channel model satisfies the update condition. The second communication device sends the second wireless channel model and the second discriminator to the first communication device. The first communications device receives and deploys a second wireless channel model and a second discriminator.
情况4:第一无线信道模型部署在第一通信设备侧,第二通信设备判别模型劣化,并指示第一通信设备执行无线信道模型的更新。Situation 4: The first wireless channel model is deployed on the side of the first communication device, and the second communication device determines that the model is degraded, and instructs the first communication device to update the wireless channel model.
第二通信设备向第一通信设备发送第二信道信息,第一通信设备接收第二通信设备发送的第二信道信息。第一通信设备通过第一无线信道模型基于第二信道信息得到第一信道信息。第二通信设备也通过第一无线信道模型基于第二信道信息得到第一信道信息。第二通信设备在通过第一判别器基于第一信道信息确定第一无线信道模型满足更新条件的情况下,向第一通信设备发送更新指示。第一通信设备接收更新指示,更新第一无线信道模型得到第二无线信道模型,第一通信设备向第二通信设备发送第二无线信道模型和第二判别器。The second communication device sends the second channel information to the first communication device, and the first communication device receives the second channel information sent by the second communication device. The first communication device obtains the first channel information based on the second channel information through the first wireless channel model. The second communication device also obtains the first channel information based on the second channel information through the first wireless channel model. The second communication device sends an update instruction to the first communication device when it is determined by the first discriminator based on the first channel information that the first wireless channel model satisfies the update condition. The first communication device receives the update instruction, updates the first wireless channel model to obtain the second wireless channel model, and sends the second wireless channel model and the second discriminator to the second communication device.
示例性的,针对第一无线信道模型包括第一编码器和第一解码器时,给出四种第一通信设备和第二通信设备的分工情况的列举。Exemplarily, when the first wireless channel model includes the first encoder and the first decoder, four cases of division of labor between the first communication device and the second communication device are enumerated.
情况5:第一解码器部署在第一通信设备侧,第一编码器部署在第二通信设备侧。第一通信设备判别模型劣化,并执行无线信道模型的更新。Case 5: the first decoder is deployed on the side of the first communication device, and the first encoder is deployed on the side of the second communication device. The first communication device determines that the model is degraded, and performs an update of the wireless channel model.
第二通信设备通过第一编码器基于第二信道信息得到第一编码结果,第二通信设备向第一通信设备发送第一编码结果,第一通信设备接收第二通信设备发送的第一编码结果。第一通信设备通过第一解码器基于第一编码结果得到第一信道信息。第一通信设备在通过第一判别器基于第一信道信息确定第一无线信道模型满足更新条件的情况下,更新第一无线信道模型得到第二无线信道模型。第一通信设备向第二通信设备发送第二编码器。第二通信设备接收第一通信设备发送的第二编码器。The second communication device obtains the first encoding result based on the second channel information through the first encoder, the second communication device sends the first encoding result to the first communication device, and the first communication device receives the first encoding result sent by the second communication device . The first communication device obtains the first channel information based on the first encoding result through the first decoder. The first communication device updates the first wireless channel model to obtain the second wireless channel model when it is determined by the first discriminator based on the first channel information that the first wireless channel model satisfies the update condition. The first communications device sends the second encoder to the second communications device. The second communication device receives the second encoder sent by the first communication device.
情况6:第一解码器部署在第一通信设备侧,第一编码器部署在第二通信设备侧。第一通信设备判别模型劣化,指示第二通信设备执行无线信道模型的更新。Case 6: The first decoder is deployed on the side of the first communication device, and the first encoder is deployed on the side of the second communication device. The first communication device determines that the discriminant model is degraded, and instructs the second communication device to update the wireless channel model.
第二通信设备通过第一编码器基于第二信道信息得到第一编码结果,第二通信设备向第一通信设备发送第一编码结果,第一通信设备接收第二通信设备发送的第一编码结果。第一通信设备通过第一解码器基于第一编码结果得到第一信道信息。第一通信设备在通过第一判别器基于第一信道信息确定第一无线信道模型满足更新条件的情况下,向第二通信设备发送更新指示。第二通信设备接收更相信指示,更新第一无线信道模型得到第二无线信道模型。第二通信设备向第一通信设备发送第二解码器和第二判别器。第一通信设备接收第二通信设备发送的第二解码器和第二判别器。The second communication device obtains the first encoding result based on the second channel information through the first encoder, the second communication device sends the first encoding result to the first communication device, and the first communication device receives the first encoding result sent by the second communication device . The first communication device obtains the first channel information based on the first encoding result through the first decoder. The first communication device sends an update instruction to the second communication device when it is determined by the first discriminator based on the first channel information that the first wireless channel model satisfies the update condition. The second communication device receives the more reliable indication, and updates the first wireless channel model to obtain the second wireless channel model. The second communication device sends the second decoder and the second discriminator to the first communication device. The first communication device receives the second decoder and the second discriminator sent by the second communication device.
情况7:第一解码器部署在第一通信设备侧,第一编码器部署在第二通信设备侧。第二通信设备判别模型劣化,并执行无线信道模型的更新。Case 7: The first decoder is deployed on the side of the first communication device, and the first encoder is deployed on the side of the second communication device. The second communication device discriminates that the model is degraded, and performs an update of the wireless channel model.
第二通信设备通过第一编码器基于第二信道信息得到第一编码结果,第二通信设备向第一通信设备发送第一编码结果,第一通信设备接收第二通信设备发送的第一编码结果。第一通信设备通过第一解码器基于第一编码结果得到第一信道信息。第二通信设备也通过第一解码器基于第一编码结果得到第一信道信息。第二通信设备在通过第一判别器基于第一信道信息确定第一无线信道模型满足更新条件的情况下,更新第一无线信道模型得到第二无线信道模型。第二通信设备向第一通信设备发送第二解码器和第二判别器。第一通信设备接收第二通信设备发送的第二解码器和第二判别器。The second communication device obtains the first encoding result based on the second channel information through the first encoder, the second communication device sends the first encoding result to the first communication device, and the first communication device receives the first encoding result sent by the second communication device . The first communication device obtains the first channel information based on the first encoding result through the first decoder. The second communication device also obtains the first channel information based on the first encoding result through the first decoder. The second communication device updates the first wireless channel model to obtain the second wireless channel model when it is determined by the first discriminator based on the first channel information that the first wireless channel model satisfies the update condition. The second communication device sends the second decoder and the second discriminator to the first communication device. The first communication device receives the second decoder and the second discriminator sent by the second communication device.
情况8:第一解码器部署在第一通信设备侧,第一编码器部署在第二通信设备侧。第二通信设备判别模型劣化,并指示第一通信设备执行无线信道模型的更新。Case 8: The first decoder is deployed on the side of the first communication device, and the first encoder is deployed on the side of the second communication device. The second communication device determines that the model is degraded, and instructs the first communication device to perform an update of the wireless channel model.
第二通信设备通过第一编码器基于第二信道信息得到第一编码结果,第二通信设备向第一通信设备发送第一编码结果,第一通信设备接收第二通信设备发送的第一编码结果。第一通信设备通过第一解码器基于第一编码结果得到第一信道信息。第二通信设备也通过第一解码器基于第一编码结果得到第一信道信息。第二通信设备在通过第一判别器基于第一信道信息确定第一无线信道模型满足更新条件的情况下,向第一通信设备发送更新指示。第一通信设备接收更相信指示,更新第一无线信道模型得到第二无线信道模型。第一通信设备向第二通信设备发送第二编码器、第二解码器和第二判别器。第二通信设备接收第二通信设备发送的第二编码器、第二解码器和第二判别器。The second communication device obtains the first encoding result based on the second channel information through the first encoder, the second communication device sends the first encoding result to the first communication device, and the first communication device receives the first encoding result sent by the second communication device . The first communication device obtains the first channel information based on the first encoding result through the first decoder. The second communication device also obtains the first channel information based on the first encoding result through the first decoder. The second communication device sends an update instruction to the first communication device when it is determined by the first discriminator based on the first channel information that the first wireless channel model satisfies the update condition. The first communication device receives the more reliable indication, and updates the first wireless channel model to obtain the second wireless channel model. The first communication device sends the second encoder, the second decoder and the second discriminator to the second communication device. The second communication device receives the second encoder, the second decoder, and the second discriminator sent by the second communication device.
其中,情况5至情况8中的第一通信设备可以替换为第二通信设备,同时第二通信设备替换为第一通信设备,也即,第一解码器部署在第二通信设备侧,第一编码器部署在第一通信设备侧。如此,还可以得到4种情况,本实施例对这四种情况不再详细说明。Wherein, the first communication device in case 5 to case 8 can be replaced by the second communication device, and at the same time the second communication device is replaced by the first communication device, that is, the first decoder is deployed on the side of the second communication device, and the first The encoder is deployed on the side of the first communication device. In this way, four situations can also be obtained, which will not be described in detail in this embodiment.
综上所述,本实施例提供的方法,在训练无线信道模型时,将无线信道模型作为生成器与判别器组成生成对抗网络,进行对抗训练,得到训练好的无线信道模型和判别器。在应用阶段,使用无线信道模型执行无线信道相关的业务,并使用判别器评价无线信道模型输出的数据,根据判别器的评价结果来确定无线信道模型的优劣。当根据评价结果确定无线信道模型劣化时,对无线信道模型进行更新。该方法可以利用训练结算训练得到的判别器,在应用阶段对模型的输出数据进行评价,从而实时感知模型输出数据的可靠性,在模型输出数据不可靠时及时对模型进行更新。To sum up, in the method provided in this embodiment, when training the wireless channel model, the wireless channel model is used as a generator and a discriminator to form an adversarial network, and the adversarial training is performed to obtain a trained wireless channel model and a discriminator. In the application stage, use the wireless channel model to perform wireless channel-related services, use the discriminator to evaluate the output data of the wireless channel model, and determine the quality of the wireless channel model according to the evaluation result of the discriminator. When it is determined that the wireless channel model is degraded according to the evaluation result, the wireless channel model is updated. This method can use the discriminator obtained from the training settlement training to evaluate the output data of the model in the application stage, so as to perceive the reliability of the model output data in real time, and update the model in time when the model output data is unreliable.
图10示出了本申请一个示例性实施例提供的无线信道模型的更新方法的架构图。本实施例以该架构应用于图1所示的终端10或网络设备20中来举例说明。Fig. 10 shows a structure diagram of a method for updating a wireless channel model provided by an exemplary embodiment of the present application. This embodiment is described by taking this architecture applied to the terminal 10 or the network device 20 shown in FIG. 1 as an example.
该架构由5个主要工作模块组成:数据收集模块501、离线联合训练模块502、在线模型推理模块503、在线判别器推理模块504、业务应用模块505。The architecture consists of five main working modules: data collection module 501 , offline joint training module 502 , online model reasoning module 503 , online discriminator reasoning module 504 , and business application module 505 .
数据收集模块501:为数据平台,实现数据过滤、数据结构化等数据预处理工作,并向离线联合训练模块502和在线模型推理模块503提供训练数据和推理数据。Data collection module 501: as a data platform, it implements data preprocessing such as data filtering and data structuring, and provides training data and inference data to the offline joint training module 502 and the online model reasoning module 503.
离线联合训练模块502:在训练数据的驱动下,联合训练第一无线信道模型和第一判别器,借鉴生成对抗网络的训练方式。通过交替训练两个模型的模型参数,直到模型收敛,完成模型训练,向在线模型推理模块503部署训练好的第一无线信道模型,向在线判别器推理模块504部署训练好的第一判别器。第一判别器对第一无线信道模型的输出起监督作用,例如监督第一无线信道模型输出是否符合真实数据分布特点,如果符合则第一判别器输出1,不符合则第一判别器输出0。示例性的,根据不同业务的不同规则,判别器的输出(评价结果)也可能是介于0~1之间的一个判别指标。Offline joint training module 502: Driven by the training data, jointly train the first wireless channel model and the first discriminator, using the training method of the generative adversarial network for reference. By alternately training the model parameters of the two models until the models converge, the model training is completed, the trained first wireless channel model is deployed to the online model reasoning module 503 , and the trained first discriminator is deployed to the online discriminator reasoning module 504 . The first discriminator plays a supervisory role on the output of the first wireless channel model, such as supervising whether the output of the first wireless channel model conforms to the distribution characteristics of the real data. If it is consistent, the first discriminator outputs 1, and if it does not conform, the first discriminator outputs 0. . Exemplarily, according to different rules of different services, the output (evaluation result) of the discriminator may also be a discriminant index between 0 and 1.
在线模型推理模块503:在线模型推理模块503接收离线联合训练模块502发送的第一无线信道模型后,完成模型部署。离线联合训练模块502开始介入业务,将推理数据(第二信道信息)输入第一无线信道模型,输出业务应用模块505所需的模块输出(第一信道信息)。将模块输出提供给业务应用模块505执行业务决策。Online model reasoning module 503: After receiving the first wireless channel model sent by the offline joint training module 502, the online model reasoning module 503 completes model deployment. The offline joint training module 502 starts to intervene in the service, inputs the reasoning data (second channel information) into the first wireless channel model, and outputs the module output (first channel information) required by the service application module 505 . The module output is provided to the business application module 505 to perform business decisions.
在线判别器推理模块504:在线判别器推理模块504接收离线联合训练模块502发送的第一判别器后,完成第一判别器部署。在线判别器推理模块504开始介入业务,将第一无线信道模型输出的模块输出输入第一判别器,第一判别器对模型输出做判别推理,得到判别器输出(评价结果)。向业务应用模块505发送判别器输出,以辅助执行业务决策。Online discriminator reasoning module 504: After receiving the first discriminator sent by the offline joint training module 502, the online discriminator reasoning module 504 completes the deployment of the first discriminator. The online discriminator reasoning module 504 starts to intervene in the business, and inputs the module output output by the first wireless channel model into the first discriminator, and the first discriminator performs discriminant reasoning on the model output to obtain the discriminator output (evaluation result). The discriminator output is sent to the business application module 505 to assist in the execution of business decisions.
示例性的,该架构中还可以包括:判别器模型在线训练模块:在线判别器推理模块504开始介入业务时,判别器模型在线训练模块同时在线预更新第一判别器的模型参数,缓存更新后得到的第二判别器。第二判别器不用于执行判别任务(不用于判别第一无线信道模型输出的第一信道信息)。第二判别器仅用于记忆新的推理数据(第二信道信息)的分布,实时学习新的推理数据分布的特点,始终保存最适合当前数据分布的模型参数。当评价结果满足更新条件时,判别器模型在线训练模块向离线联合训练模块502发送第二判别器,驱动离线联合训练模块502基于第二判别器训练第一无线信道模型,得到更符合当前数据分布的第二无线信道模型。Exemplarily, the architecture may also include: discriminator model online training module: when the online discriminator reasoning module 504 starts to intervene in the business, the discriminator model online training module simultaneously pre-updates the model parameters of the first discriminator online, and after the cache update The resulting second discriminator. The second discriminator is not used to perform a discrimination task (not used to discriminate the first channel information output by the first wireless channel model). The second discriminator is only used to memorize the distribution of new reasoning data (second channel information), learn the characteristics of the new reasoning data distribution in real time, and always save the model parameters most suitable for the current data distribution. When the evaluation result satisfies the update condition, the discriminator model online training module sends the second discriminator to the offline joint training module 502, and drives the offline joint training module 502 to train the first wireless channel model based on the second discriminator to obtain a model that is more in line with the current data distribution. The second wireless channel model of .
业务应用模块505:业务应用模块505接收在线模型推理模块503发送的模块输出,接收在线判别器 推理模块504输出的判别器输出。若判别器输出满足预设条件,则业务应用模块505将直接使用第一无线信道模型的输出作为业务决策依据。若判别器输出不满足预设条件,则业务应用模块505将不会使用第一无线信道模型的输出,示例性的,在这种情况下,可以采用其他策略流程(不使用神经网络模型的策略流程)执行业务。Business application module 505: The business application module 505 receives the module output sent by the online model reasoning module 503, and receives the discriminator output output by the online discriminator reasoning module 504. If the output of the discriminator satisfies the preset condition, the service application module 505 will directly use the output of the first wireless channel model as a basis for service decision. If the output of the discriminator does not meet the preset condition, the service application module 505 will not use the output of the first wireless channel model. Exemplarily, in this case, other strategy processes (the strategy of not using the neural network model process) to execute the business.
示例性的,本申请实施例示例性地提供了多种更新条件的设置方案,本申请实施例还提供了多种更新方式。Exemplarily, the embodiment of the present application exemplarily provides multiple update condition setting schemes, and the embodiment of the present application also provides multiple update methods.
图11示出了本申请一个示例性实施例提供的无线信道模型的更新方法的流程图。本实施例以该方法应用于第一通信设备来举例说明,第一通信设备可以是图1所示的终端10或网络设备20。该方法包括:Fig. 11 shows a flowchart of a method for updating a wireless channel model provided by an exemplary embodiment of the present application. This embodiment is described by taking the method applied to a first communication device as an example, and the first communication device may be the terminal 10 or the network device 20 shown in FIG. 1 . The method includes:
步骤401:使用训练样本对抗训练初始无线信道模型和初始判别器得到第一无线信道模型和第一判别器。Step 401: Use training samples to confront training an initial wireless channel model and an initial discriminator to obtain a first wireless channel model and a first discriminator.
在应用第一无线信道模型和第一判别器之前,先训练得到第一无线信道模型和第一判别器。Before applying the first wireless channel model and the first discriminator, the first wireless channel model and the first discriminator are obtained through training.
示例性的,确定无线信道模型和判别器的模型结构,初始化模型参数得到初始无线信道模型和初始判别器。针对初始无线信道模型和初始判别器组成的生成对抗网络,使用训练样本对抗训练初始无线信道模型和初始判别器,得到第一无线信道模型和第一判别器。Exemplarily, the wireless channel model and the model structure of the discriminator are determined, and model parameters are initialized to obtain an initial wireless channel model and an initial discriminator. Aiming at the generative adversarial network composed of the initial wireless channel model and the initial discriminator, the initial wireless channel model and the initial discriminator are trained against the training samples to obtain the first wireless channel model and the first discriminator.
示例性的,步骤401还可以由第二通信设备执行,当步骤401由第二通信设备执行时,第一通信设备接收第二通信设备发送的第一无线信道模型和第一判别器。Exemplarily, step 401 may also be performed by the second communication device, and when step 401 is performed by the second communication device, the first communication device receives the first wireless channel model and the first discriminator sent by the second communication device.
步骤410:通过第一无线信道模型得到第一信道信息,第一无线信道模型是作为生成器与第一判别器对抗训练得到的神经网络模型。Step 410: Obtain first channel information through a first wireless channel model, where the first wireless channel model is a neural network model that is trained as a generator against a first discriminator.
示例性的,通过第一无线信道模型基于第二信道信息得到第一信道信息。将一个第二信道信息输入第一无线信道模型得到一个信道数据。步骤410中的第一信道信息可以是指一个信道数据,也可以是指第一无线信道模型基于多个第二信道信息输出的多个信道数据。Exemplarily, the first channel information is obtained based on the second channel information through the first wireless channel model. Inputting a second channel information into the first wireless channel model to obtain a channel data. The first channel information in step 410 may refer to one channel data, or may refer to multiple channel data output by the first wireless channel model based on multiple second channel information.
步骤420:通过第一判别器评价第一信道信息得到第一概率值。Step 420: Evaluate the first channel information by a first discriminator to obtain a first probability value.
示例性的,第一判别器根据输入的一个信道数据输出一个概率值。示例性的,第一无线信道模型每输出一个信道模型,将其输入第一判别器输出一个概率值。第一概率值包括与第一信道信息中每个信道数据一一对应的至少一个概率值。Exemplarily, the first discriminator outputs a probability value according to an input channel data. Exemplarily, each time a channel model is output by the first wireless channel model, it is input into the first discriminator to output a probability value. The first probability value includes at least one probability value corresponding to each channel data in the first channel information.
概率值用于表示第一信道信息是真实数据或非真实数据。或,概率值包括0或1,0表示非真实数据,1表示真实数据。或,概率值为0到1的数值(小数或整数),接近0表示接近非真实数据,接近1表示接近真实数据。The probability value is used to indicate whether the first channel information is real data or non-real data. Alternatively, the probability value includes 0 or 1, where 0 indicates non-real data and 1 indicates real data. Or, a numerical value (decimal or integer) with a probability value from 0 to 1, close to 0 means close to non-real data, close to 1 means close to real data.
其中,真实数据是指通过实际采集或其他数据获取方式获得的数据。非真实数据(模拟数据/生成数据)是指无线信道模型输出的数据。Among them, real data refers to data obtained through actual collection or other data acquisition methods. Non-real data (simulated data/generated data) refers to the data output by the wireless channel model.
示例性的,第一判别器可以包括多个判别器,使用多个判别器从多个维度上对信道信息的真实性(是否为真实信道数据)进行判定。则第一概率值包括多个判别器基于第一信道信息输出的评价结果。Exemplarily, the first discriminator may include multiple discriminators, and multiple discriminators are used to determine the authenticity of the channel information (whether it is real channel data) from multiple dimensions. Then the first probability value includes evaluation results output by multiple discriminators based on the first channel information.
步骤421:基于第一信道信息和第一概率值执行无线信道业务。Step 421: Perform a wireless channel service based on the first channel information and the first probability value.
示例性的,第一信道信息和第一概率值都可以应用于执行无线信道相关的业务。相关技术中的方法,无线信道模型只应用第一信道信息执行业务决策,而本申请实施例提供的方法,还提供了第一概率值来辅助执行业务决策。Exemplarily, both the first channel information and the first probability value may be applied to perform services related to wireless channels. In the method in the related art, the radio channel model only uses the first channel information to execute service decision, while the method provided in the embodiment of the present application also provides the first probability value to assist in the execution of service decision.
第一概率值可以评价第一信道信息的可靠性。以第一信道信息是一个信道信息为例,若第一概率值为非真实数据,则说明第一信道信息不可靠,不可以使用第一信道信息执行业务决策;若第一概率值为真实数据,则说明第一信道信息可靠,可以使用第一信道信息执行业务决策。The first probability value can evaluate the reliability of the first channel information. Taking the first channel information as an example, if the first probability value is not real data, it means that the first channel information is unreliable, and the first channel information cannot be used to perform business decisions; if the first probability value is real data , it indicates that the first channel information is reliable, and the first channel information can be used to perform service decisions.
进一步的,若第一概率值是0到1的数值,则应用第一信道信息执行信道业务时,也可以将第一概率值作为第一信道信息的权重系数,从而在将第一信道信息应用于实际业务时,体现出第一信道信息的可靠性。Further, if the first probability value is a value from 0 to 1, when applying the first channel information to perform channel services, the first probability value may also be used as the weight coefficient of the first channel information, so that when the first channel information is applied to In actual business, it reflects the reliability of the first channel information.
示例性的,本申请实施例提供的方法,在将第一无线信道模型输出的信道信息应用于信道业务之前,就可以使用判别器来实时感知第一无线信道模型是否劣化,不需要在将信道信息应用于信道业务后由业务反馈性能指标。该方法可以在应用信道信息前及时感知第一无线信道模型的劣化,以便及时更新无线信道模型,防止模型劣化影响业务处理。Exemplarily, in the method provided by the embodiment of this application, before the channel information output by the first wireless channel model is applied to the channel service, the discriminator can be used to sense whether the first wireless channel model is degraded in real time, without the need to apply the channel information After the information is applied to the channel service, the service feeds back the performance index. The method can perceive the degradation of the first wireless channel model in time before applying the channel information, so as to update the wireless channel model in time and prevent the model degradation from affecting service processing.
步骤430:在基于第一概率值确定第一无线信道模型满足更新条件的情况下,更新第一无线信道模型得到第二无线信道模型。Step 430: In a case where it is determined based on the first probability value that the first wireless channel model satisfies the update condition, update the first wireless channel model to obtain a second wireless channel model.
示例性的,对于是否满足更新条件的判定,可以是实时判定,也可以是在输出指定数量的信道信息后进行判定,还可以是周期性地进行判定。即,可以在第一无线信道模型每输出一个信道信息后,都基于该信道信息的概率值,确定是否满足更新条件,以便实时更新无线信道模型。也可以在第一无线信道模型输出指定数量的信道信息后,基于该指定数量的信道信息的概率值,确定是否满足更新条件。还可以周期性 地根据在该周期内得到的概率值,确定是否满足更新条件。Exemplarily, the determination of whether the update condition is satisfied may be performed in real time, after outputting a specified amount of channel information, or periodically. That is, after each piece of channel information is output by the first wireless channel model, it may be determined whether the update condition is satisfied based on the probability value of the channel information, so as to update the wireless channel model in real time. It may also be determined whether the update condition is satisfied based on the probability value of the specified amount of channel information after the first wireless channel model outputs the specified amount of channel information. It is also possible to periodically determine whether the update condition is met according to the probability value obtained in the period.
示例性的,第一信道信息包括多个信道信息,第一概率值包括与多个信道信息一一对应的多个概率值;判别器用于判别输入数据为真实信道数据(真实数据)或非真实信道数据(非真实数据)。Exemplarily, the first channel information includes a plurality of channel information, and the first probability value includes a plurality of probability values corresponding to the plurality of channel information; the discriminator is used to distinguish whether the input data is real channel data (real data) or non-real Channel data (not real data).
更新条件是基于第一概率值设置的用于评价无线信道模型是否劣化的评判标准,更新条件的设置可以是任意的,基于第一概率值来表示无线信道模型劣化程度的方式都可以用于实现更新条件的设置,本申请实施例对于更新条件的设置不加以限定。The update condition is a criterion for evaluating whether the wireless channel model is degraded based on the first probability value. The setting of the update condition can be arbitrary, and the method of expressing the degree of degradation of the wireless channel model based on the first probability value can be used to realize The setting of the update condition, the embodiment of the present application does not limit the setting of the update condition.
示例性的,本申请实施例提供了几种可选的更新条件。更新条件包括如下4个条件中的至少一种:Exemplarily, the embodiment of this application provides several optional update conditions. Update conditions include at least one of the following four conditions:
条件1:第一概率值中概率值低于第一阈值的占比高于预设值。Condition 1: The proportion of the probability values lower than the first threshold among the first probability values is higher than the preset value.
即,第一概率值中概率值为非真实信道数据(即,概率值低于第一阈值,例如,低于0.5)的占比高于预设值。That is, the proportion of the probability value in the first probability value that is not real channel data (that is, the probability value is lower than the first threshold, for example, lower than 0.5) is higher than the preset value.
示例性的,用非真实信道数据的占比来表示无线信道模型的劣化程度,占比越高则说明无线信道模型劣化程度搞;占比越低则说明无线信道模型越好。Exemplarily, the degradation degree of the wireless channel model is represented by the ratio of the non-real channel data, the higher the ratio, the higher the degradation degree of the wireless channel model; the lower the ratio, the better the wireless channel model.
示例性的,第一无线信道模型每输出一个信道信息,都可以计算一次占比,即,第一概率值包括第一判别器最近输出的多个概率值,计算第一判别器最近输出的多个概率值中表示非真实信道数据的占比,若占比较高则说明模型劣化,需要进行模型更新。Exemplarily, each time the first wireless channel model outputs a piece of channel information, the proportion can be calculated once, that is, the first probability value includes a plurality of probability values recently output by the first discriminator, and the number of recent output by the first discriminator is calculated. A probability value represents the proportion of non-real channel data. If the proportion is high, it indicates that the model is degraded and needs to be updated.
示例性的,也可以在第一无线信道模型输出指定数量的信道信息后,计算一次占比,即,第一概率值包括指定数量的概率值。例如,每当第一无线信道模型输出100个信道信息后,基于该100个信道信息的100个概率值计算非真实信道数据的占比。Exemplarily, the proportion may also be calculated once after the first wireless channel model outputs a specified amount of channel information, that is, the first probability value includes the specified number of probability values. For example, whenever the first wireless channel model outputs 100 pieces of channel information, the proportion of non-real channel data is calculated based on 100 probability values of the 100 pieces of channel information.
示例性的,还可以周期性的计算占比,即,每隔一段时间基于第一判别器在该算时间内输出的多个概率值计算占比。即,第一概率值包括第一判别器在第一周期内输出的多个概率值,第一周期的周期时长为预设时长。例如,每隔一个小时,基于第一无线信道模型在1小时内输出的58个信道信息的58个概率值,计算其中非真实信道数据的占比。Exemplarily, the proportion can also be calculated periodically, that is, the proportion can be calculated based on multiple probability values output by the first discriminator within the calculation time at intervals. That is, the first probability value includes a plurality of probability values output by the first discriminator in the first cycle, and the cycle duration of the first cycle is a preset duration. For example, every hour, based on 58 probability values of 58 channel information output by the first wireless channel model within 1 hour, the proportion of non-real channel data is calculated.
条件2:第一概率值中连续的x个概率值低于第二阈值。Condition 2: x consecutive probability values among the first probability values are lower than the second threshold.
即,第一概率值中连续的x个概率值为非真实信道数据(即,连续x个概率值低于第二阈值,例如,低于0.5),x为正整数。That is, the x consecutive probability values in the first probability value are not real channel data (that is, the x consecutive probability values are lower than the second threshold, for example, lower than 0.5), and x is a positive integer.
条件2还可以表述为:第一判别器最近输出的x个概率值为非真实信道数据。或,多个概率值中存在连续的x个概率值为非真实信道数据,其中,多个概率值为第一判别器在第一周期内输出的,或,多个概率值为第一判别器输出的指定数量的概率值。Condition 2 can also be expressed as: the latest x probability values output by the first discriminator are not real channel data. Or, among the plurality of probability values, there are consecutive x probability values that are not real channel data, wherein the plurality of probability values are output by the first discriminator in the first cycle, or, the plurality of probability values are output by the first discriminator Outputs the specified number of probability values.
示例性的,第一概率值包括第一判别器最近输出的x个概率值。每当第一判别器输出一个概率值,则判定第一判别器最近输出的x个概率值是否都为非真实信道数据,若都为非真实信道数据则第一概率值满足更新条件,第一无线信道模型需要进行更新。Exemplarily, the first probability value includes the latest x probability values output by the first discriminator. Whenever the first discriminator outputs a probability value, it is determined whether the x probability values recently output by the first discriminator are all unreal channel data. If they are all unreal channel data, the first probability value satisfies the update condition, and the first The radio channel model needs to be updated.
示例性的,第一概率值也可以是第一判别器在第一周期内输出的N个概率值,N为大于x的整数,判定第一概率值中是否存在连续的x个概率值果都为非真实信道数据。Exemplarily, the first probability value may also be N probability values output by the first discriminator in the first period, N is an integer greater than x, and it is determined whether there are consecutive x probability values in the first probability value. is non-true channel data.
示例性的,第一概率值也可以是第一判别器连续输出的M个概率值,M为预设值,M为大于x的整数,判定第一概率值中是否存在连续的x个概率值都为非真实信道数据。Exemplarily, the first probability value may also be M probability values continuously output by the first discriminator, M is a preset value, M is an integer greater than x, and it is determined whether there are consecutive x probability values in the first probability value All are non-true channel data.
条件3:根据第一概率值得到的概率值分布情况满足第一条件。Condition 3: The probability value distribution obtained according to the first probability value satisfies the first condition.
示例性的,概率值可以是0到1的数值(包括小数和整数),则可以根据多个概率值概率值分布图/曲线,基于概率值分布图/曲线上的参数来判定第一无线信道模型是否劣化。例如,根据概率值分布图/曲线中的曲率、斜率、变化率、聚类情况等来判定第一无线信道模型输出的信道信息的可靠性,进而判断第一无线信道模型是否劣化。Exemplarily, the probability value may be a value from 0 to 1 (including decimals and integers), then the first wireless channel may be determined based on parameters on the probability value distribution graph/curve according to multiple probability value distribution graphs/curves Whether the model is degraded. For example, the reliability of the channel information output by the first wireless channel model is determined according to the curvature, slope, rate of change, clustering, etc. in the probability value distribution diagram/curve, and then whether the first wireless channel model is degraded.
概率值分布情况满足第一条件是指,概率值分布情况表现出信道数据不可信/信道数据为非真实信道数据。The fact that the probability value distribution satisfies the first condition means that the probability value distribution shows that the channel data is not credible/the channel data is unreal channel data.
条件4:基于第一概率值计算得到的评价数值达到第三阈值。Condition 4: The evaluation value calculated based on the first probability value reaches the third threshold.
示例性的,还可以基于第一概率值计算得到其他用于评价模型劣化程度的评价数值,用该评价数值是否达到第三阈值来判定是否满足更新条件。Exemplarily, other evaluation values for evaluating the degree of model degradation may also be calculated based on the first probability value, and whether the evaluation value reaches a third threshold is used to determine whether the update condition is satisfied.
上述四个条件可以独立设置为更新条件,也可以任意个组合设置为更新条件。The above four conditions can be set as update conditions independently, or can be set as update conditions in any combination.
例如,条件1和条件2可以一起设置为更新条件。即,在第一概率值中概率值低于第一阈值的占比高于预设值,且,第一概率值中连续的x个概率值低于第二阈值的情况下,确定第一概率值满足更新条件,触发更新第一无线信道模型。For example, Condition 1 and Condition 2 can be set together as an update condition. That is, when the proportion of the probability values lower than the first threshold value in the first probability value is higher than the preset value, and the consecutive x probability values in the first probability value are lower than the second threshold value, the first probability value is determined The value satisfies the update condition, triggering the update of the first wireless channel model.
示例性的,第一判别器还可以包括多个判别器,例如第一判别器包括两个判别器。不同判别器从不同的维度上评价第一信道信息的真实性。示例性的,第一判别器包括至少两个判别器,第一概率值包括至少两个判别器分别输出的至少两个子概率值。则在基于至少两个子概率值中的至少一个子概率值确定第一无 线信道模型满足更新条件的情况下,触发更新第一无线信道模型。Exemplarily, the first discriminator may further include multiple discriminators, for example, the first discriminator includes two discriminators. Different discriminators evaluate the authenticity of the first channel information from different dimensions. Exemplarily, the first discriminator includes at least two discriminators, and the first probability value includes at least two sub-probability values respectively output by the at least two discriminators. Then, when it is determined based on at least one sub-probability value in the at least two sub-probability values that the first wireless channel model satisfies the update condition, the update of the first wireless channel model is triggered.
在基于第一概率值确定第一无线信道模型满足更新条件后,触发更新第一无线信道模型。本申请实施例提供了两种更新模型的方式:After it is determined based on the first probability value that the first wireless channel model satisfies the update condition, an update of the first wireless channel model is triggered. The embodiment of this application provides two ways to update the model:
方式1:获取第二信道数据,将第二信道数据(正样本)作为训练样本,实时训练更新第一判别器得到第二判别器;在基于第一概率值确定第一无线信道模型满足更新条件后,使用第二判别器与第一无线信道模型组成生成对抗网络,对抗训练第一无线信道模型得到第二无线信道模型。Method 1: Obtain the second channel data, use the second channel data (positive sample) as a training sample, train and update the first discriminator in real time to obtain the second discriminator; determine that the first wireless channel model satisfies the update condition based on the first probability value Finally, the second discriminator and the first wireless channel model are used to form a generated confrontation network, and the first wireless channel model is trained against to obtain the second wireless channel model.
在应用第一无线信道模型得到第一信道信息后,使用第一判别器评价第一信道信息得到第一概率值,同时,使用第二信道信息对第一判别器进行训练得到第二判别器。即,不断用第二信道信息来训练第一判别器,使第二判别器始终能够学习到最新的业务数据的数据分布特征。After applying the first wireless channel model to obtain the first channel information, use the first discriminator to evaluate the first channel information to obtain the first probability value, and at the same time, use the second channel information to train the first discriminator to obtain the second discriminator. That is, the first discriminator is continuously trained with the second channel information, so that the second discriminator can always learn the latest data distribution characteristics of service data.
需要注意的是,训练得到的第二判别器并不会用于评价第一信道信息,在训练得到第二判别器后,还是使用第一判别器对第一信道信息进行评价,即,第一判别器用于评价第一信道信息,第二判别器用于更新第一无线信道模型。第二判别器用于与第一无线信道模型组成生成对抗网络,对抗训练第一无线信道模型得到第二无线信道模型。相当于,第一判别器是实时进行更新的,而第一无线信道模型是在基于第一概率值确定第一无线信道模型满足更新条件后更新的,在对第一无线信道模型进行更新时固定第二判别器的模型参数不变,即,不需要再训练第二判别器,直接利用第二判别器对抗训练第一无线信道模型即可。It should be noted that the trained second discriminator will not be used to evaluate the first channel information. After the second discriminator is trained, the first discriminator is still used to evaluate the first channel information, that is, the first The discriminator is used to evaluate the first channel information, and the second discriminator is used to update the first wireless channel model. The second discriminator is used to form a generation confrontation network with the first wireless channel model, and the first wireless channel model is trained against to obtain the second wireless channel model. It is equivalent to that the first discriminator is updated in real time, and the first wireless channel model is updated after it is determined that the first wireless channel model satisfies the update condition based on the first probability value, and is fixed when the first wireless channel model is updated. The model parameters of the second discriminator remain unchanged, that is, there is no need to retrain the second discriminator, and the first wireless channel model can be directly trained against the second discriminator.
即,使用第二信道信息迭代训练第一判别器,得到第二判别器,第二判别器用于更新第一无线信道模型。That is, the first discriminator is iteratively trained using the second channel information to obtain a second discriminator, and the second discriminator is used to update the first wireless channel model.
示例性的,第二信道信息包括输入第一无线信道模型的m个信道信息,m为正整数;使用第j个训练样本训练第一判别器,第j个训练样本包括第j个信道信息和第j个标签,第j个标签将第j个信道信息标注为真实信道数据,j为不大于m的正整数;令j=j+1,重复执行上述步骤迭代训练第一判别器得到第二判别器。Exemplarily, the second channel information includes m channel information input to the first wireless channel model, where m is a positive integer; the jth training sample is used to train the first discriminator, and the jth training sample includes the jth channel information and The jth label, the jth label marks the jth channel information as real channel data, j is a positive integer not greater than m; let j=j+1, repeat the above steps to iteratively train the first discriminator to obtain the second Discriminator.
训练第一无线信道模型的方式为:对于由第一无线信道模型和第二判别器组成的第二生成对抗网络,保持第二判别器的参数不变,使用训练样本对抗训练第一无线信道模型得到第二无线信道模型。The way to train the first wireless channel model is: for the second generative confrontation network composed of the first wireless channel model and the second discriminator, keep the parameters of the second discriminator unchanged, and use the training samples to train the first wireless channel model A second wireless channel model is obtained.
示例性的,实时迭代训练第一判别器不仅会用到第二信道信息(正样本),还会用到第一无线信道模型输出的第一信道信息作为负样本,实时迭代训练第一判别器得到第二判别器。Exemplarily, the real-time iterative training of the first discriminator will not only use the second channel information (positive samples), but also use the first channel information output by the first wireless channel model as a negative sample, and iteratively train the first discriminator in real time Get the second discriminator.
方式2:在基于第一概率值确定第一无线信道模型满足更新条件后,获取第二信道信息;针对第一判别器与第一无线信道模型组成的生成对抗网络,将第二信道信息和第一信道信息作为训练样本,进行对抗训练,得到第二无线信道模型和第二判别器。Method 2: After determining that the first wireless channel model satisfies the update condition based on the first probability value, obtain the second channel information; for the generative confrontation network composed of the first discriminator and the first wireless channel model, combine the second channel information and the first A channel information is used as a training sample, and confrontation training is performed to obtain a second wireless channel model and a second discriminator.
方式2则不需要实时训练第一判别器,只需要在满足更新条件后,统一基于第一无线信道模型和第一判别器进行交替进行对抗训练,得到训练好的第二无线信道模型和第二判别器。Method 2 does not need to train the first discriminator in real time, but only needs to conduct alternate confrontation training based on the first wireless channel model and the first discriminator after the update conditions are met, and obtain the trained second wireless channel model and the second discriminator. Discriminator.
步骤440:通过第二无线信道模型得到第三信道信息。Step 440: Obtain third channel information through the second wireless channel model.
示例性的,在更新第一无线信道模型得到第二无线信道模型,以及第二无线信道模型对应的第二判别器后,继续应用第二无线信道模型基于第四信道信息得到第三信道信息,使用第二判别器评价第三信道信息得到第二概率值,基于第三信道信息和第三概率值执行无线信道相关业务。Exemplarily, after updating the first wireless channel model to obtain the second wireless channel model and the second discriminator corresponding to the second wireless channel model, continue to apply the second wireless channel model to obtain the third channel information based on the fourth channel information, A second discriminator is used to evaluate the third channel information to obtain a second probability value, and wireless channel related services are performed based on the third channel information and the third probability value.
步骤450:通过第二判别器评价第三信道信息得到第二概率值。Step 450: Evaluate the third channel information by the second discriminator to obtain a second probability value.
步骤460:在基于第二概率值确定第二无线信道模型满足更新条件的情况下,更新第二无线信道模型。Step 460: When it is determined based on the second probability value that the second radio channel model satisfies the update condition, update the second radio channel model.
继续利用第二判别器输出的第二概率值,监测第二无线信道模型是否需要进行更新。Continue to use the second probability value output by the second discriminator to monitor whether the second wireless channel model needs to be updated.
综上所述,本实施例提供的方法,提供了多种设置更新条件的方法,保证模型更新的实时性。并且,对于无线信道模型的更新,可以采用实时更新判别器的方式,使判别器实时学习数据分布变化情况,在对无线信道模型更新时,使用实时更新得到的判别器对抗训练无线信道模型,使无线信道模型也可能够学习到数据分布变化情况,进而输出更真实的信道信息。该方法可以在触发模型更新的同时,向模型反馈数据布局变化趋势,进而引导模型更新方向,提高模型更新效率。To sum up, the method provided in this embodiment provides multiple methods for setting update conditions to ensure real-time model update. Moreover, for the update of the wireless channel model, the method of updating the discriminator in real time can be adopted, so that the discriminator can learn the change of data distribution in real time. The wireless channel model may also be able to learn the changes in data distribution, and then output more realistic channel information. This method can feed back the change trend of data layout to the model while triggering the model update, thereby guiding the direction of model update and improving the efficiency of model update.
示例性的,给出一种使用本申请提供的无线信道模型的更新方法,更新信道估计模型的示例性实施例。Exemplarily, an exemplary embodiment of updating a channel estimation model by using the method for updating a wireless channel model provided in this application is given.
图12示出了本申请一个示例性实施例提供的无线信道模型的更新方法的流程图。本实施例以该方法应用于第一通信设备中来举例说明,第一通信设备可以是图1所示的终端10或网络设备20。该方法包括:Fig. 12 shows a flowchart of a method for updating a wireless channel model provided by an exemplary embodiment of the present application. This embodiment is described by taking the method applied to a first communication device as an example, and the first communication device may be the terminal 10 or the network device 20 shown in FIG. 1 . The method includes:
步骤510:通过第一信道估计模型得到第一信道估计结果,第一信道估计模型是作为生成器与第一判别器对抗训练得到的神经网络模型。Step 510: Obtain a first channel estimation result through the first channel estimation model, where the first channel estimation model is a neural network model obtained by adversarial training as a generator and a first discriminator.
示例性的,该方法由终端设备或网络设备的在线环境模块执行。第一信道估计模型和第一判别器是在离线环境中对抗训练得到的。Exemplarily, the method is executed by an online environment module of a terminal device or a network device. The first channel estimation model and the first discriminator are obtained by adversarial training in an offline environment.
第二通信设备可以是终端设备或网络设备,第二通信设备向第一通信设备发送第二信道信息,第一通信设备通过第一信道估计模型基于第二信道信息得到第一信道估计结果。第二信道信息可以是接收参考信号和真实参考信号。或,第二信道信息可以是已估计的信道数据。The second communication device may be a terminal device or a network device, the second communication device sends the second channel information to the first communication device, and the first communication device obtains the first channel estimation result based on the second channel information through the first channel estimation model. The second channel information may be a received reference signal and a real reference signal. Or, the second channel information may be estimated channel data.
示例性的,第二信道信息是输入第一信道估计模型的数据,第一信道估计结果是第一信道估计模型输出的数据,将第二信道信息输入第一信道估计模型,第一信道估计模型基于输入的第二信道信息输出第一信道估计结果。Exemplarily, the second channel information is the data input into the first channel estimation model, the first channel estimation result is the data output by the first channel estimation model, and the second channel information is input into the first channel estimation model, and the first channel estimation model A first channel estimation result is output based on the input second channel information.
示例性的,如图13所示,在执行步骤510之前,执行步骤601,联合训练第一信道估计模型和第一判别器。示例性的,如图14所示,信道估计模型701和判别器702组成生成对抗网络。在训练阶段信道估计模型701的输入是导频数据集(接收参考信号+真实参考信号),或,输入是已估计的信道数据(例如,图4中所示的信道估计步骤中得到的已估计的信道数据),输出是信道数据集(信道估计结果),将NMSE函数作为损失函数,度量信道估计结果和真实信道数据的误差(损失值),基于损失值更新信道估计模型501的模型参数。Exemplarily, as shown in FIG. 13 , before step 510 is performed, step 601 is performed to jointly train the first channel estimation model and the first discriminator. Exemplarily, as shown in FIG. 14 , a channel estimation model 701 and a discriminator 702 form a generative confrontation network. In the training phase, the input of the channel estimation model 701 is the pilot data set (received reference signal + real reference signal), or the input is the estimated channel data (for example, the estimated channel data obtained in the channel estimation step shown in Fig. 4 channel data), the output is a channel data set (channel estimation result), the NMSE function is used as a loss function, the error (loss value) between the channel estimation result and the real channel data is measured, and the model parameters of the channel estimation model 501 are updated based on the loss value.
第一判别器702的输入是信道数据集(信道估计结果),其中信道估计模型输出的信道估计结果标记0,实际的真实信道数据标记1。判别器可以理解成一个二元分类器,判别器的损失函数可以是交叉熵损失函数。判别器输出0到1之间的一个数,若判别器完全分辨不出输入的是信道估计结果还是真实信道数据,那么输出0.5。The input of the first discriminator 702 is a channel data set (channel estimation result), wherein the channel estimation result output by the channel estimation model is marked 0, and the actual real channel data is marked 1. The discriminator can be understood as a binary classifier, and the loss function of the discriminator can be a cross-entropy loss function. The discriminator outputs a number between 0 and 1. If the discriminator cannot tell whether the input is the channel estimation result or the real channel data, then output 0.5.
第一信道估计模型和第一判别器的训练过程可以是:首先,初始化信道估计模型和判别器;将训练样本输入信道估计模型得到信道估计结果(标记0);将信道估计模型的模型参数固定不变,用信道估计结果(负样本)和真实信道数据(正样本)训练判别器,让判别器清晰的分辨出真实信道数据和信道估计结果(二元分类),得到训练好的判别器,此时判别器可以分清真实信道数据和信道估计结果;然后,解除信道估计模型的模型参数固定,固定判别器的模型参数,用训练样本训练信道估计模型直到模型收敛,此时信道估计模型生成的信道估计结果是更加真实而且符合真实信道数据的潜在特点的信道估计结果。The training process of the first channel estimation model and the first discriminator can be: first, initialize the channel estimation model and the discriminator; Input the training sample into the channel estimation model to obtain the channel estimation result (mark 0); the model parameters of the channel estimation model are fixed Unchanged, use the channel estimation results (negative samples) and real channel data (positive samples) to train the discriminator, let the discriminator clearly distinguish the real channel data and channel estimation results (binary classification), and get the trained discriminator, At this time, the discriminator can distinguish between real channel data and channel estimation results; then, the model parameters of the channel estimation model are released, the model parameters of the discriminator are fixed, and the channel estimation model is trained with training samples until the model converges. At this time, the channel estimation model generates The channel estimation result is more realistic and conforms to the potential characteristics of real channel data.
需要注意的是,上述的生成对抗网络的训练步骤仅仅是选用了一种比较基本且常用的训练方式,还可以采用其他生成对抗网络的训练方法对其进行对抗训练。It should be noted that the training steps of the above-mentioned generative adversarial network only select a relatively basic and commonly used training method, and other training methods of generative adversarial networks can also be used for adversarial training.
在训练得到第一信道估计模型和第一判别器后,执行步骤602,离线环境模块向在线环境模块发送第一信道估计模型和第一判别器。步骤603,在线环境模块部署第一信道估计模型和第一判别器。随后执行步骤604,使用第一信道估计模型和第一判别器在线推理(输出信道估计结果和评价结果)。After the first channel estimation model and the first discriminator are obtained through training, step 602 is executed, and the offline environment module sends the first channel estimation model and the first discriminator to the online environment module. Step 603, the online environment module deploys the first channel estimation model and the first discriminator. Then step 604 is executed, using the first channel estimation model and the first discriminator for online reasoning (outputting channel estimation results and evaluation results).
步骤520:通过第一判别器评价第一信道估计结果得到第一概率值。Step 520: Evaluate the first channel estimation result by the first discriminator to obtain the first probability value.
如图13所示,在推理过程中,执行步骤605,第一判别器始终在线学习新的数据分布,持续更新第一判别器的模型参数得到第二判别器。As shown in FIG. 13 , during the reasoning process, step 605 is executed, the first discriminator always learns new data distribution online, and continuously updates the model parameters of the first discriminator to obtain the second discriminator.
步骤530:在基于第一概率值确定第一信道估计模型满足更新条件的情况下,更新第一信道估计模型得到第二信道估计模型。Step 530: In a case where it is determined based on the first probability value that the first channel estimation model satisfies the update condition, update the first channel estimation model to obtain a second channel estimation model.
在基于第一概率值确定第一无线信道模型满足更新条件的情况下,向离线环境模块发送第一更新指示;接收离线环境模块发送的第二信道估计模型,第二信道估计模型是离线更新第一信道估计模型得到的。When determining that the first wireless channel model satisfies the update condition based on the first probability value, send the first update instruction to the offline environment module; receive the second channel estimation model sent by the offline environment module, and the second channel estimation model is the offline update first A channel estimation model is obtained.
第一更新指示还包括第二判别器,第二判别器是使用第二信道信息在线实时训练第一判别器得到的;第二信道估计模型是根据第二判别器离线对抗训练第一信道估计模型得到的。The first update instruction also includes a second discriminator, the second discriminator is obtained by using the second channel information to train the first discriminator online in real time; the second channel estimation model is based on the second discriminator offline confrontation training of the first channel estimation model owned.
如图13所示,步骤606,当环境变化等因素导致数据分布变化时,可以利用第一判别器输出的评价结果直接感知到当前第一信道估计模型的劣化程度。示例性的,统计判别器输出的负面判别结果(例如判别器输出小于0.8的次数);当最近一段时间窗内,负面判别结果概率超过某一门限(例如30%),意味着当前信道估计模型已经劣化,为了避免模型持续失效,带来进一步的性能恶化,需要触发模型更新。步骤607在劣化程度超出预算门限时,触发模型更新。步骤608,向离线环境模块发送第二判别器。步骤609,离线环境模块使用第二判别器更新第一信道估计模型得到第二信道估计模型,此时第二信道估计模型生成的信道估计结果更加符合最近一段真实信道数据的潜在分布特点。步骤610,离线环境模块向在线环境模块发送第二信道估计模型。在线环境模块删除第一判别器,使用第二判别器代替第一判别器,使用第二信道估计模型代替第一信道估计模型,继续执行推理过程。As shown in FIG. 13 , in step 606 , when factors such as environmental changes lead to changes in data distribution, the evaluation results output by the first discriminator can be used to directly perceive the degree of degradation of the current first channel estimation model. Exemplarily, the negative discriminant results output by the statistical discriminator (for example, the number of times the discriminator output is less than 0.8); when the probability of a negative discriminative result exceeds a certain threshold (for example, 30%) within a recent period of time window, it means that the current channel estimation model has been degraded, in order to avoid continuous failure of the model and further performance degradation, it is necessary to trigger a model update. Step 607 triggers model update when the degree of degradation exceeds the budget threshold. Step 608, sending the second discriminator to the offline environment module. In step 609, the offline environment module uses the second discriminator to update the first channel estimation model to obtain a second channel estimation model. At this time, the channel estimation result generated by the second channel estimation model is more in line with the latent distribution characteristics of the latest real channel data. Step 610, the offline environment module sends the second channel estimation model to the online environment module. The online environment module deletes the first discriminator, uses the second discriminator to replace the first discriminator, uses the second channel estimation model to replace the first channel estimation model, and continues to execute the reasoning process.
综上所述,本实施例提供的方法,在训练信道估计模型时,将信道估计模型作为生成器与判别器组成生成对抗网络,进行对抗训练,得到训练好的信道估计模型和判别器。在应用阶段,使用信道估计模型执行信道估计相关的业务,并使用判别器评价信道估计模型输出的数据,根据判别器的评价结果来确定信道估计模型的优劣。当根据评价结果确定信道估计模型劣化时,对信道估计模型进行更新。该方法可以利用训练结算训练得到的判别器,在应用阶段对模型的输出数据进行评价,从而实时感知模型输出数据的可靠性,在模型输出数据不可靠时及时对模型进行更新。To sum up, in the method provided by this embodiment, when training the channel estimation model, the channel estimation model is used as a generator and a discriminator to form an adversarial network, and adversarial training is performed to obtain a trained channel estimation model and a discriminator. In the application phase, the channel estimation model is used to perform channel estimation-related services, and the discriminator is used to evaluate the output data of the channel estimation model, and the quality of the channel estimation model is determined according to the evaluation result of the discriminator. When it is determined that the channel estimation model is degraded according to the evaluation result, the channel estimation model is updated. This method can use the discriminator obtained from the training settlement training to evaluate the output data of the model in the application stage, so as to perceive the reliability of the model output data in real time, and update the model in time when the model output data is unreliable.
示例性的,给出一种使用本申请提供的无线信道模型的更新方法,更新CSI自编码模型的示例性实施例。Exemplarily, an exemplary embodiment of updating a CSI self-encoding model by using the wireless channel model updating method provided in this application is given.
图15示出了本申请一个示例性实施例提供的无线信道模型的更新方法的流程图。本实施例以该方法应用于第一通信设备中来举例说明,第一通信设备可以是图1所示的网络设备20(接入网设备)。该方法 包括:Fig. 15 shows a flowchart of a method for updating a wireless channel model provided by an exemplary embodiment of the present application. This embodiment is described by taking the method applied to the first communication device as an example, and the first communication device may be the network device 20 (access network device) shown in FIG. 1 . The method includes:
步骤710:通过第一CSI自编码模型得到第一CSI恢复结果,第一CSI自编码模型是作为生成器与第一判别器对抗训练得到的神经网络模型。Step 710: Obtain a first CSI recovery result by using the first CSI autoencoder model, which is a neural network model obtained by training as a generator against the first discriminator.
第一CSI自编码模型包括第一CSI编码器和第一CSI解码器。第一CSI自编码模型用于在终端设备侧对CSI反馈信息进行编码压缩,将压缩后的编码结果发送至接入网设备侧,在接入网设备侧进行解码得到恢复出来的CSI反馈信息。第一CSI自编码模型用于编码并恢复CSI反馈信息,即,在理想情况下第一CSI自编码模型的输入CSI反馈信息和输出CSI反馈信息(第一CSI恢复结果)是相同的。The first CSI autoencoder model includes a first CSI encoder and a first CSI decoder. The first CSI self-encoding model is used to encode and compress the CSI feedback information on the terminal equipment side, send the compressed encoding result to the access network equipment side, and decode the recovered CSI feedback information on the access network equipment side. The first CSI autoencoder model is used to encode and restore CSI feedback information, that is, ideally, the input CSI feedback information and output CSI feedback information (first CSI restoration result) of the first CSI autoencoder model are the same.
即,第二信道信息是CSI反馈信息/CSI数据。第二信道信息是输入第一CSI自编码模型的数据,第一CSI恢复结果是第一CSI自编码模型的输出数据。将第二信道信息输入第一CSI编码器,得到第一编码结果,将第一编码结果输入第一CSI解码器,得到第一CSI恢复结果。That is, the second channel information is CSI feedback information/CSI data. The second channel information is data input to the first CSI auto-encoding model, and the first CSI restoration result is output data of the first CSI auto-encoding model. The second channel information is input to the first CSI encoder to obtain a first encoding result, and the first encoding result is input to the first CSI decoder to obtain a first CSI recovery result.
在步骤510之前,接收终端设备发送的第一CSI解码器和第一判别器,第一CSI解码器和第一判别器是终端设备训练得到的。接收终端设备发送的第一编码结果,第一编码结果是终端设备通过第一CSI编码器对真实CSI数据(第二信道信息)编码得到的。步骤510包括:通过第一CSI解码器对第一编码结果进行解码得到第一CSI恢复结果(第一信道信息)。Before step 510, the first CSI decoder and the first discriminator sent by the terminal device are received, and the first CSI decoder and the first discriminator are obtained through training by the terminal device. The first encoding result sent by the terminal device is received, where the first encoding result is obtained by the terminal device encoding real CSI data (second channel information) by using a first CSI encoder. Step 510 includes: decoding the first coding result by a first CSI decoder to obtain a first CSI restoration result (first channel information).
如图16所示,CSI自编码模型包括CSI编码器703和CSI解码器704。CSI编码器部署在终端设备侧,CSI解码器部署在接入网设备侧。CSI编码器将输入的CSI数据编码得到第一编码结果,终端设备将第一编码结果发送给接入网设备,接入网设备利用CSI解码器对第一编码结果进行解码得到第一CSI恢复结果。将第一CSI恢复结果输入第一判别器705得到评价结果。判别器用于区分CSI恢复结果和真实CSI数据。As shown in FIG. 16 , the CSI self-encoding model includes a CSI encoder 703 and a CSI decoder 704 . The CSI encoder is deployed on the terminal device side, and the CSI decoder is deployed on the access network device side. The CSI encoder encodes the input CSI data to obtain the first encoding result, the terminal device sends the first encoding result to the access network device, and the access network device uses the CSI decoder to decode the first encoding result to obtain the first CSI recovery result . Input the first CSI recovery result into the first discriminator 705 to obtain the evaluation result. The discriminator is used to distinguish the CSI recovery results from real CSI data.
步骤720:通过第一判别器评价第一CSI恢复结果得到第一概率值。Step 720: Evaluate the first CSI restoration result by the first discriminator to obtain the first probability value.
步骤730:在基于第一概率值确定第一CSI自编码模型满足更新条件的情况下,更新第一CSI自编码模型得到第二CSI自编码模型。Step 730: When it is determined based on the first probability value that the first CSI autoencoder model satisfies the update condition, update the first CSI autoencoder model to obtain a second CSI autoencoder model.
第二CSI自编码模型包括第二CSI编码器和第二CSI解码器。The second CSI autoencoder model includes a second CSI encoder and a second CSI decoder.
在基于第一概率值确定第一CSI自编码模型满足更新条件的情况下,向终端设备发送第二更新指示,第二更新指示用于指示终端设备更新第一CSI自编码模型和第一判别器;接收终端设备发送的第二CSI解码器和第二判别器,第二CSI解码器和第二判别器是对抗训练第一CSI自编码模型和第一判别器得到的。When it is determined based on the first probability value that the first CSI autoencoding model satisfies the update condition, a second update instruction is sent to the terminal device, and the second update instruction is used to instruct the terminal device to update the first CSI autoencoder model and the first discriminator ; receiving the second CSI decoder and the second discriminator sent by the terminal device, where the second CSI decoder and the second discriminator are obtained by adversarial training of the first CSI autoencoder model and the first discriminator.
如图17所示,步骤801,终端设备训练第一CSI自编码模型和第一判别器。步骤802,终端设备向接入网设备发送第一CSI解码器和第一判别器。步骤803,接入网设备部署第一CSI解码器和第一判别器。步骤804,终端设备收集保存最近一段时间内的CSI数据。步骤805,接入网设备使用第一CSI解码器和第一判别器在线推理输出第一CSI恢复结果和第一概率值。步骤806,接入网设备基于第一判别器输出的概率值统计劣化程度。步骤807,在劣化程度超出预算门限时,触发模型更新。当第一CSI解码器的输出不像真实CSI数据时,第一判别器可以做出即时反馈。结合第一判别器的判别结果,接入网设备可以立刻判断第一CSI解码器解析的CSI恢复结果是合格的CSI数据。为下一步的决策提供参考。当环境变化等因素导致数据分布变化时,接入网设备可以利用第一判别器可以直接感知到CSI自编码模型的劣化程度,触发模型更新。步骤808,接入网设备向终端设备发送第二更新指示。步骤809,终端设备利用步骤804中保存的CSI数据更新第一CSI自编码模型和第一判别器得到第二CSI自编码模型和第二判别器。步骤810,终端设备向接入网设备发送第二CSI解码器和第二判别器。As shown in FIG. 17 , in step 801, the terminal device trains a first CSI autoencoder model and a first discriminator. Step 802, the terminal device sends the first CSI decoder and the first discriminator to the access network device. Step 803, the access network device deploys the first CSI decoder and the first discriminator. Step 804, the terminal device collects and saves the CSI data within a recent period of time. Step 805, the access network device uses the first CSI decoder and the first discriminator to infer online and output the first CSI restoration result and the first probability value. In step 806, the access network device calculates the degree of degradation based on the probability value output by the first discriminator. Step 807, when the degree of degradation exceeds the budget threshold, trigger model update. When the output of the first CSI decoder does not look like real CSI data, the first discriminator can make immediate feedback. Combined with the discrimination result of the first discriminator, the access network device can immediately determine that the CSI restoration result analyzed by the first CSI decoder is qualified CSI data. Provide a reference for the next step of decision-making. When factors such as environmental changes lead to changes in data distribution, the access network device can use the first discriminator to directly perceive the degree of degradation of the CSI self-encoding model, and trigger a model update. Step 808, the access network device sends a second update instruction to the terminal device. In step 809, the terminal device uses the CSI data saved in step 804 to update the first CSI autoencoder model and the first discriminator to obtain the second CSI autoencoder model and the second discriminator. Step 810, the terminal device sends the second CSI decoder and the second discriminator to the access network device.
综上所述,本实施例提供的方法,在训练CSI自编码模型时,将CSI自编码模型作为生成器与判别器组成生成对抗网络,进行对抗训练,得到训练好的CSI自编码模型和判别器。在应用阶段,使用CSI自编码模型执行信道估计相关的业务,并使用判别器评价CSI自编码模型输出的数据,根据判别器的评价结果来确定CSI自编码模型的优劣。当根据评价结果确定CSI自编码模型劣化时,对CSI自编码模型进行更新。该方法可以利用训练结算训练得到的判别器,在应用阶段对模型的输出数据进行评价,从而实时感知模型输出数据的可靠性,在模型输出数据不可靠时及时对模型进行更新。To sum up, the method provided in this embodiment, when training the CSI autoencoder model, uses the CSI autoencoder model as a generator and a discriminator to form an adversarial network, conducts confrontation training, and obtains the trained CSI autoencoder model and discriminator device. In the application phase, the CSI self-encoding model is used to perform channel estimation-related services, and the discriminator is used to evaluate the output data of the CSI self-encoding model, and the quality of the CSI self-encoding model is determined according to the evaluation result of the discriminator. When it is determined that the CSI autoencoder model is degraded according to the evaluation result, the CSI autoencoder model is updated. This method can use the discriminator obtained from the training settlement training to evaluate the output data of the model in the application stage, so as to perceive the reliability of the model output data in real time, and update the model in time when the model output data is unreliable.
示例性的,给出一种使用本申请提供的无线信道模型的更新方法,更新CSI自编码模型的示例性实施例。Exemplarily, an exemplary embodiment of updating a CSI self-encoding model by using the wireless channel model updating method provided in this application is given.
图18示出了本申请一个示例性实施例提供的无线信道模型的更新方法的流程图。本实施例以该方法应用于第一通信设备中来举例说明,第一通信设备可以是图1所示的网络设备20(接入网设备)。该方法包括:Fig. 18 shows a flowchart of a method for updating a wireless channel model provided by an exemplary embodiment of the present application. This embodiment is described by taking the method applied to the first communication device as an example, and the first communication device may be the network device 20 (access network device) shown in FIG. 1 . The method includes:
步骤910:通过第一CSI预测自编码模型得到第一CSI预测结果,第一CSI预测自编码模型是作为生成器与第一判别器对抗训练得到的神经网络模型。Step 910: Obtain a first CSI prediction result through the first CSI prediction autoencoder model, which is a neural network model obtained by training as a generator against the first discriminator.
第一CSI预测自编码模型包括第一CSI预测编码器和第一CSI预测解码器。示例性的,第一CSI预测自编码模型用于在终端设备侧,通过CSI预测编码器基于连续的N个历史周期CSI进行编码得到第二编码结果,将第二编码结果发送给接入网设备侧,接入网设备侧通过CSI预测解码器对第二编码结果进行解码, 得到第一CSI预测结果。第一CSI预测结果是基于N个历史周期CSI得到的预测CSI序列。即,参考图5中的图示,第一CSI预测自编码模型用于基于N个历史周期CSI预测得到未来时频上的CSI序列。The first CSI predictive self-encoding model includes a first CSI predictive encoder and a first CSI predictive decoder. Exemplarily, the first CSI predictive self-encoding model is used on the side of the terminal device to obtain a second encoding result by encoding the CSI predictive encoder based on the CSI of consecutive N historical periods, and send the second encoding result to the access network device On the access network device side, the second encoding result is decoded by a CSI prediction decoder to obtain the first CSI prediction result. The first CSI prediction result is a predicted CSI sequence obtained based on N historical period CSIs. That is, referring to the illustration in FIG. 5 , the first CSI prediction self-encoding model is used to obtain a future time-frequency CSI sequence based on CSI prediction of N historical periods.
第一CSI预测结果包括CSI序列。The first CSI prediction result includes a CSI sequence.
第二信道信息包括N个历史周期CSI。将第二信道信息输入胡第一CSI预测编码器,得到第二编码结果,将第二编码结果输入第一CSI预测解码器,得到第一CSI预测结果(第一信道信息)。The second channel information includes N historical period CSIs. Input the second channel information into the first CSI prediction encoder to obtain the second encoding result, and input the second encoding result into the first CSI prediction decoder to obtain the first CSI prediction result (first channel information).
在步骤910之前,接入网设备接收终端设备发送的第一CSI预测编码器和第一判别器,第一CSI预测编码器和第一判别器是终端设备训练得到的。接入网设备接收终端设备发送的第二编码结果,第二编码结果是终端设备通过第一CSI预测编码器对CSI序列编码得到的。步骤910包括:通过第一CSI预测解码器对第二编码结果进行解码得到第一CSI预测结果。Before step 910, the access network device receives the first CSI predictive coder and the first discriminator sent by the terminal device, and the first CSI predictive coder and the first discriminator are trained by the terminal device. The access network device receives the second encoding result sent by the terminal device, where the second encoding result is obtained by the terminal device by encoding the CSI sequence through the first CSI predictive encoder. Step 910 includes: decoding the second encoding result by a first CSI prediction decoder to obtain a first CSI prediction result.
如图19所示,CSI预测自编码模型包括CSI预测编码器706和CSI预测解码器707。CSI预测编码器部署在终端设备侧,CSI预测解码器部署在接入网设备侧。CSI预测编码器将输入的连续N个历史周期CSI编码得到第二编码结果,终端设备将第二编码结果发送给接入网设备,接入网设备利用CSI预测解码器对第二编码结果进行解码得到第一CSI预测结果(预测CSI序列)。将第一CSI预测结果输入时空关系判别器705和时序关系判别器709得到评价结果。时空关系判别器705和时序关系判别器709用于分别在时空和时序维度上评价CSI预测结果的真实性。As shown in FIG. 19 , the CSI prediction self-encoding model includes a CSI prediction encoder 706 and a CSI prediction decoder 707 . The CSI predictive encoder is deployed on the terminal device side, and the CSI predictive decoder is deployed on the access network device side. The CSI predictive encoder encodes the input CSI for N consecutive historical periods to obtain the second encoding result, the terminal device sends the second encoding result to the access network device, and the access network device uses the CSI predictive decoder to decode the second encoding result Obtain the first CSI prediction result (predicted CSI sequence). Input the first CSI prediction result into the spatio-temporal relationship discriminator 705 and the temporal relationship discriminator 709 to obtain the evaluation result. The spatiotemporal relationship discriminator 705 and the temporal relationship discriminator 709 are used to evaluate the authenticity of the CSI prediction results in the spatiotemporal and temporal dimensions, respectively.
步骤920:通过第一判别器评价第一CSI预测结果得到第一概率值。Step 920: Evaluate the first CSI prediction result by the first discriminator to obtain the first probability value.
示例性的,第一判别器包括第一时序关系判别器和第一时空关系判别器,第一概率值包括第一时序概率值和第一时空概率值。Exemplarily, the first discriminator includes a first temporal relationship discriminator and a first spatiotemporal relationship discriminator, and the first probability value includes a first temporal probability value and a first spatiotemporal probability value.
时序关系判别器用于从CSI预测结果的时间维度上评价CSI预测结果的真实性。时空关系判别器用于从CSI预测结果的空间维度上评价CSI预测结果的真实性。The temporal relationship discriminator is used to evaluate the authenticity of the CSI prediction results from the time dimension of the CSI prediction results. The spatio-temporal relationship discriminator is used to evaluate the authenticity of the CSI prediction results from the spatial dimension of the CSI prediction results.
时序关系判别器是学习实际CSI序列的时间相关性特点,并判别CSI预测自编码模型输出的预测CSI序列(CSI预测结果)时间关系特性是否符合实际CSI序列的时间相关性特点。时空关系判别器是学习CSI本身的空间特性,关注空间特征的提取,以此判别CSI预测自编码模型输出的CSI(CSI预测结果)空间分布特征是否符合实际CSI数据的空间分布特点。The timing relationship discriminator is to learn the temporal correlation characteristics of the actual CSI sequence, and judge whether the temporal correlation characteristics of the predicted CSI sequence (CSI prediction result) output by the CSI prediction self-encoding model conform to the temporal correlation characteristics of the actual CSI sequence. The spatio-temporal relationship discriminator learns the spatial characteristics of CSI itself and pays attention to the extraction of spatial features, so as to judge whether the spatial distribution characteristics of CSI (CSI prediction results) output by the CSI prediction autoencoder model conform to the spatial distribution characteristics of actual CSI data.
步骤930:在基于第一概率值确定第一CSI预测自编码模型满足更新条件的情况下,更新第一CSI预测自编码模型得到第二CSI预测自编码模型。Step 930: When it is determined based on the first probability value that the first CSI predictive auto-encoding model satisfies the update condition, update the first CSI predictive auto-encoder model to obtain a second CSI predictive auto-encoder model.
第二CSI预测自编码模型包括第二CSI预测编码器和第二CSI预测解码器。The second CSI predictive autoencoder model includes a second CSI predictive encoder and a second CSI predictive decoder.
在第一时序概率值和第一时空概率值中的至少一个满足更新条件的情况下,更新第一CSI预测自编码模型得到第二CSI预测自编码模型。In a case where at least one of the first time-series probability value and the first space-time probability value satisfies the update condition, the first CSI predictive auto-encoding model is updated to obtain the second CSI predictive auto-encoder model.
示例性的,接入网设备接收终端设备发送的第二CSI预测解码器和第二判别器;其中,第二CSI预测解码器和第二判别器是终端设备,在第一时序概率值和第一时空概率值中的至少一个满足更新条件的情况下,更新第一CSI预测自编码模型和第一判别器得到的。Exemplarily, the access network device receives the second CSI predictive decoder and the second discriminator sent by the terminal device; wherein, the second CSI predictive decoder and the second discriminator are terminal devices, and the first timing probability value and the second discriminator In a case where at least one of the spatiotemporal probability values satisfies the update condition, the first CSI prediction obtained from the self-encoding model and the first discriminator is updated.
如图20所示,步骤901,终端设备训练第一CSI预测自编码模型和第一判别器。步骤902,终端设备向接入网设备发送第一CSI预测解码器和第一判别器。步骤903,接入网设备部署第一CSI预测解码器和第一判别器。步骤904,终端设备在线学习,利用应用过程(推理过程)中所产生的CSI预测结果和数据库中的真实CSI序列持续更新第一判别器的模型参数得到第二判别器。步骤905,接入网设备使用第一CSI预测解码器和第一判别器在线推理输出第一CSI预测结果和第一概率值。步骤906,终端设备基于第一判别器输出的概率值统计劣化程度。步骤907,在劣化程度超出预算门限时,触发模型更新。当CSI预测自编码模型输出的预测CSI序列不像一个真实CSI序列时,第一判别器可以做出即时反馈。结合第一判别器的概率值,接入网设备可以立刻判断CSI预测解码器解析的CSI预测结果是否可靠。为下一步的决策提供参考。第一判别器可以始终在终端设备侧学习新的数据分布,预更新模型参数。当环境变化等因素导致数据分布变化时,终端设备利用第一始判别器可以直接感知到CSI预测自编码模型的劣化程度,触发更新。终端设备利用实时更新得到的第二判别器对抗训练第一CSI预测自编码模型得到第二CSI预测自编码模型。步骤908,终端设备向接入网设备发送第二CSI预测解码器和第二判别器。As shown in FIG. 20, in step 901, the terminal device trains a first CSI prediction autoencoder model and a first discriminator. Step 902, the terminal device sends the first CSI prediction decoder and the first discriminator to the access network device. Step 903, the access network device deploys a first CSI prediction decoder and a first discriminator. Step 904, the terminal device learns online, continuously updates the model parameters of the first discriminator by using the CSI prediction results generated in the application process (reasoning process) and the real CSI sequence in the database to obtain the second discriminator. Step 905, the access network device uses the first CSI prediction decoder and the first discriminator to infer online and output the first CSI prediction result and the first probability value. Step 906, the terminal device calculates the degree of degradation based on the probability value output by the first discriminator. Step 907, when the degree of degradation exceeds the budget threshold, trigger model update. When the predicted CSI sequence output by the CSI prediction autoencoder model does not resemble a real CSI sequence, the first discriminator can make immediate feedback. Combined with the probability value of the first discriminator, the access network device can immediately determine whether the CSI prediction result analyzed by the CSI prediction decoder is reliable. Provide a reference for the next step of decision-making. The first discriminator can always learn new data distribution on the terminal device side and pre-update model parameters. When factors such as environmental changes lead to changes in data distribution, the terminal device can directly perceive the degree of degradation of the CSI prediction self-encoding model by using the first initial discriminator, and trigger an update. The terminal device adversarially trains the first CSI predictive autoencoder model by using the second discriminator obtained by updating in real time to obtain the second CSI predictive autoencoder model. Step 908, the terminal device sends the second CSI prediction decoder and the second discriminator to the access network device.
综上所述,本实施例提供的方法,在训练CSI预测自编码模型时,将CSI预测自编码模型作为生成器与判别器组成生成对抗网络,进行对抗训练,得到训练好的CSI预测自编码模型和判别器。在应用阶段,使用CSI预测自编码模型执行信道估计相关的业务,并使用判别器评价CSI预测自编码模型输出的数据,根据判别器的概率值来确定CSI预测自编码模型的优劣。当根据概率值确定CSI预测自编码模型劣化时,对CSI预测自编码模型进行更新。该方法可以利用训练结算训练得到的判别器,在应用阶段对模型的输出数据进行评价,从而实时感知模型输出数据的可靠性,在模型输出数据不可靠时及时对模型进行更新。To sum up, the method provided in this embodiment, when training the CSI predictive autoencoder model, uses the CSI predictive autoencoder model as a generator and a discriminator to form an adversarial network, conducts confrontation training, and obtains a trained CSI predictive autoencoder Model and Discriminator. In the application stage, use the CSI predictive self-encoding model to perform channel estimation-related services, and use the discriminator to evaluate the output data of the CSI predictive self-encoding model, and determine the pros and cons of the CSI predictive self-encoding model according to the probability value of the discriminator. When it is determined that the CSI prediction self-encoding model is degraded according to the probability value, the CSI prediction self-encoding model is updated. This method can use the discriminator obtained from the training settlement training to evaluate the output data of the model in the application stage, so as to perceive the reliability of the model output data in real time, and update the model in time when the model output data is unreliable.
图22示出了本申请一个示例性实施例提供的一种第一通信装置的框图,所述装置包括:Fig. 22 shows a block diagram of a first communication device provided by an exemplary embodiment of the present application, and the device includes:
第一模型模块21,用于通过第一无线信道模型得到第一信道信息,所述第一无线信道模型为第一生成 对抗网络中的生成器;The first model module 21 is used to obtain the first channel information by the first wireless channel model, and the first wireless channel model is a generator in the first generated confrontation network;
第一更新模块22,用于在通过第一判别器基于所述第一信道信息确定所述第一无线信道模型满足更新条件的情况下,触发更新所述第一无线信道模型,所述第一判别器为所述第一生成对抗网络中的判别器。The first update module 22 is configured to trigger an update of the first wireless channel model when the first discriminator determines that the first wireless channel model satisfies an update condition based on the first channel information, and the first The discriminator is a discriminator in the first generation confrontation network.
在一个可选的实施例中,所述第一更新模块22包括:In an optional embodiment, the first updating module 22 includes:
第一发送子模块24,用于向第二通信装置发送第一更新指示,所述第一更新指示用于触发所述第二通信装置更新所述第一无线信道模型得到第二无线信道模型;The first sending submodule 24 is configured to send a first update instruction to a second communication device, where the first update instruction is used to trigger the second communication device to update the first wireless channel model to obtain a second wireless channel model;
第一接收子模块23,用于接收所述第二通信装置发送的所述第二无线信道模型。The first receiving submodule 23 is configured to receive the second wireless channel model sent by the second communication device.
在一个可选的实施例中,所述第一模型模块21,用于通过所述第一无线信道模型基于第二信道信息得到所述第一信道信息;In an optional embodiment, the first model module 21 is configured to obtain the first channel information based on the second channel information through the first wireless channel model;
所述装置还包括:The device also includes:
第一实时训练模块28,用于将所述第二信道信息作为训练样本实时训练更新所述第一判别器得到第二判别器;The first real-time training module 28 is used to use the second channel information as a training sample to train and update the first discriminator in real time to obtain a second discriminator;
其中,所述第一更新指示包括所述第二判别器;所述第二判别器用于与所述第一无线信道模型组成第二生成对抗网络训练更新所述第一无线信道模型。Wherein, the first update instruction includes the second discriminator; the second discriminator is used to form a second generative adversarial network with the first wireless channel model to train and update the first wireless channel model.
在一个可选的实施例中,所述第一更新模块22,用于更新所述第一无线信道模型得到第二无线信道模型。In an optional embodiment, the first updating module 22 is configured to update the first wireless channel model to obtain a second wireless channel model.
在一个可选的实施例中,所述第一模型模块21,用于通过所述第一无线信道模型基于第二信道信息得到所述第一信道信息;In an optional embodiment, the first model module 21 is configured to obtain the first channel information based on the second channel information through the first wireless channel model;
所述装置还包括:The device also includes:
第一实时训练模块28,用于将所述第二信道信息作为训练样本实时训练更新所述第一判别器得到第二判别器;The first real-time training module 28 is used to use the second channel information as a training sample to train and update the first discriminator in real time to obtain a second discriminator;
所述第一更新模块22,用于基于所述第二判别器和所述第一无线信道模型组成的第二生成对抗网络,训练更新所述第一无线信道模型得到所述第二无线信道模型。The first updating module 22 is configured to train and update the first wireless channel model to obtain the second wireless channel model based on the second generative confrontation network composed of the second discriminator and the first wireless channel model .
在一个可选的实施例中,所述第一更新模块22,用于基于所述第一判别器和所述第一无线信道模型组成的所述第一生成对抗网络,训练更新所述第一判别器和所述第一无线信道模型得到第二判别器和所述第二无线信道模型。In an optional embodiment, the first updating module 22 is configured to train and update the first GAN based on the first GAN composed of the first discriminator and the first wireless channel model. A discriminator and the first radio channel model obtain a second discriminator and the second radio channel model.
在一个可选的实施例中,所述装置还包括:In an optional embodiment, the device also includes:
第一接收模块27,用于接收第二通信装置发送的第二信道信息;The first receiving module 27 is configured to receive the second channel information sent by the second communication device;
所述第一模型模块21,用于通过所述第一无线信道模型基于所述第二信道信息得到所述第一信道信息。The first model module 21 is configured to obtain the first channel information based on the second channel information through the first wireless channel model.
在一个可选的实施例中,所述第一无线信道模型包括第一编码器和第一解码器,所述第一解码器部署在所述第一通信装置侧,所述第一编码器部署在所述第二通信装置侧;所述第二无线信道模型包括第二编码器和第二解码器;In an optional embodiment, the first wireless channel model includes a first encoder and a first decoder, the first decoder is deployed on the side of the first communication device, and the first encoder is deployed On the side of the second communication device; the second wireless channel model includes a second encoder and a second decoder;
所述第一接收子模块23,用于接收所述第二通信装置发送的所述第二解码器和第二判别器。The first receiving submodule 23 is configured to receive the second decoder and the second discriminator sent by the second communication device.
在一个可选的实施例中,所述装置还包括:In an optional embodiment, the device also includes:
第一接收模块27,用于接收所述第二通信装置发送的第一编码结果,所述第一编码结果是通过所述第一编码器基于第二信道信息得到的;The first receiving module 27 is configured to receive a first encoding result sent by the second communication device, where the first encoding result is obtained by the first encoder based on the second channel information;
所述第一模型模块21,用于通过所述第一解码器基于所述第一编码结果得到所述第一信道信息。The first model module 21 is configured to use the first decoder to obtain the first channel information based on the first encoding result.
在一个可选的实施例中,所述第一无线信道模型包括第一编码器和第一解码器,所述第一编码器部署在所述第一通信装置侧,所述第一解码器部署在第二通信装置侧,所述第一通信装置侧存储有所述第一解码器;所述第二无线信道模型包括第二编码器和第二解码器;In an optional embodiment, the first wireless channel model includes a first encoder and a first decoder, the first encoder is deployed on the side of the first communication device, and the first decoder is deployed On the side of the second communication device, the first decoder is stored on the side of the first communication device; the second wireless channel model includes a second encoder and a second decoder;
所述第一更新模块22包括:The first updating module 22 includes:
第一发送子模块24,用于向所述第二通信装置发送所述第二解码器和所述第二判别器。The first sending sub-module 24 is configured to send the second decoder and the second discriminator to the second communication device.
在一个可选的实施例中,In an alternative embodiment,
所述第一无线信道模型包括第一信道估计模型;The first wireless channel model includes a first channel estimation model;
所述第一信道信息包括第一信道估计结果;The first channel information includes a first channel estimation result;
所述第二信道信息包括参考信号。The second channel information includes a reference signal.
在一个可选的实施例中,In an alternative embodiment,
所述第一无线信道模型包括第一CSI自编码模型,所述第二无线信道模型包括第二CSI自编码模型;The first wireless channel model includes a first CSI self-encoding model, and the second wireless channel model includes a second CSI self-encoding model;
所述第一信道信息包括第一CSI恢复结果。The first channel information includes a first CSI recovery result.
在一个可选的实施例中,所述第一通信装置包括终端设备,所述第二通信装置包括网络设备;In an optional embodiment, the first communication device includes a terminal device, and the second communication device includes a network device;
所述第一无线信道模型包括第一CSI预测自编码模型,所述第二无线信道模型包括第二CSI预测自编 码模型;The first wireless channel model includes a first CSI predictive self-encoding model, and the second wireless channel model includes a second CSI predictive self-encoding model;
所述第一信道信息包括第一CSI预测结果。The first channel information includes a first CSI prediction result.
在一个可选的实施例中,所述更新模块,包括;In an optional embodiment, the update module includes;
第一评价子模块26,用于通过所述第一判别器评价所述第一信道信息得到第一概率值;The first evaluation submodule 26 is configured to evaluate the first channel information through the first discriminator to obtain a first probability value;
第一更新子模块25,用于在基于所述第一概率值确定所述第一无线信道模型满足更新条件的情况下,触发更新所述第一无线信道模型。The first updating submodule 25 is configured to trigger updating of the first wireless channel model when it is determined based on the first probability value that the first wireless channel model satisfies an update condition.
在一个可选的实施例中,所述第一信道信息包括多个信道信息,所述第一概率值包括与所述多个信道信息一一对应的多个概率值;In an optional embodiment, the first channel information includes a plurality of channel information, and the first probability value includes a plurality of probability values corresponding to the plurality of channel information one-to-one;
所述更新条件包括如下条件中的至少一种:The update conditions include at least one of the following conditions:
所述第一概率值中概率值低于第一阈值的占比高于预设值;The proportion of probability values lower than the first threshold among the first probability values is higher than a preset value;
所述第一概率值中连续的x个概率值低于第二阈值,x为正整数;x consecutive probability values in the first probability value are lower than the second threshold, and x is a positive integer;
根据所述第一概率值得到的概率值分布情况满足第一条件;The probability value distribution obtained according to the first probability value satisfies the first condition;
基于所述第一概率值计算得到的评价数值达到第三阈值。The evaluation value calculated based on the first probability value reaches a third threshold.
在一个可选的实施例中,所述第一判别器包括至少两个子判别器,所述第一概率值包括至少两个子概率值。In an optional embodiment, the first discriminator includes at least two sub-discriminators, and the first probability value includes at least two sub-probability values.
在一个可选的实施例中,所述第一更新子模块25,用于在基于所述至少两个子概率值中的至少一个子概率值确定所述第一无线信道模型满足更新条件的情况下,触发更新所述第一无线信道模型。In an optional embodiment, the first update submodule 25 is configured to determine that the first wireless channel model satisfies an update condition based on at least one sub-probability value of the at least two sub-probability values , triggering an update of the first wireless channel model.
图23示出了本申请一个示例性实施例提供的一种无线信道模型的更新装置的框图,该装置可以实现成为上述的第二通信装置,所述装置包括:FIG. 23 shows a block diagram of an apparatus for updating a wireless channel model provided by an exemplary embodiment of the present application. The apparatus can be realized as the above-mentioned second communication apparatus, and the apparatus includes:
第二更新模块32,用于在通过第一判别器基于第一信道信息确定第一无线信道模型满足更新条件的情况下,更新所述第一无线信道模型,所述第一信道信息是所述第一无线信道模型输出的,所述第一无线信道模型为第一生成对抗网络中的生成器,所述第一判别器为所述第一生成对抗网络中的判别器。The second update module 32 is configured to update the first wireless channel model when the first discriminator determines that the first wireless channel model satisfies the update condition based on the first channel information, the first channel information being the The output of the first wireless channel model is that the first wireless channel model is a generator in the first generative adversarial network, and the first discriminator is a discriminator in the first generative adversarial network.
在一个可选的实施例中,所述第二更新模块32,包括:In an optional embodiment, the second updating module 32 includes:
第二接收子模块33,用于接收第一通信装置发送的第一更新指示,所述第一更新指示是所述第一通信装置在通过所述第一判别器基于所述第一信道信息确定所述第一无线信道模型满足更新条件的情况下发送的;The second receiving submodule 33 is configured to receive the first update instruction sent by the first communication device, the first update instruction is determined by the first communication device based on the first channel information through the first discriminator Sent when the first wireless channel model satisfies an update condition;
第二更新子模块34,用于更新所述第一无线信道模型得到第二无线信道模型;The second updating submodule 34 is configured to update the first wireless channel model to obtain a second wireless channel model;
第二发送子模块35,用于向所述第一通信装置发送所述第二无线信道模型。The second sending submodule 35 is configured to send the second wireless channel model to the first communication device.
在一个可选的实施例中,所述第一更新指示包括第二判别器,所述第二判别器是使用第二信道信息实时训练更新所述第一判别器得到的,所述第二信道信息是所述第一无线信道模型生成所述第一信道信息时的输入信息;In an optional embodiment, the first update instruction includes a second discriminator, the second discriminator is obtained by using the second channel information to train and update the first discriminator in real time, and the second channel The information is input information when the first wireless channel model generates the first channel information;
所述第二更新子模块34,用于在所述第二判别器与所述第一无线信道模型组成的第二生成对抗网络中,训练更新所述第一无线信道模型得到所述第二无线信道模型。The second update submodule 34 is configured to train and update the first wireless channel model to obtain the second wireless channel model in the second generative confrontation network composed of the second discriminator and the first wireless channel model. channel model.
在一个可选的实施例中,所述装置还包括:In an optional embodiment, the device also includes:
第二发送模块36,用于向所述第一通信装置发送第二信道信息,所述第二信道信息是所述第一无线信道模型生成所述第一信道信息时的输入信息。The second sending module 36 is configured to send second channel information to the first communication device, where the second channel information is input information when the first wireless channel model generates the first channel information.
在一个可选的实施例中,所述第一无线信道模型包括第一编码器和第一解码器,所述第一解码器部署在所述第一通信装置侧,所述第一编码器部署在所述第二通信装置侧;所述第二无线信道模型包括第二编码器和第二解码器;In an optional embodiment, the first wireless channel model includes a first encoder and a first decoder, the first decoder is deployed on the side of the first communication device, and the first encoder is deployed On the side of the second communication device; the second wireless channel model includes a second encoder and a second decoder;
所述第二发送子模块35,用于向所述第一通信装置发送所述第二解码器和第二判别器。The second sending submodule 35 is configured to send the second decoder and the second discriminator to the first communication device.
在一个可选的实施例中,所述装置还包括:In an optional embodiment, the device also includes:
第二模型模块31,用于通过所述第一编码器基于第二信道信息得到第一编码结果,所述第二信道信息是所述第一无线信道模型生成所述第一信道信息时的输入信息;The second model module 31 is configured to use the first encoder to obtain a first encoding result based on second channel information, and the second channel information is an input when the first wireless channel model generates the first channel information information;
第二发送模块36,用于向所述第一通信装置发送所述第一编码结果,所述第一信道信息是所述第一通信装置通过所述第一解码器基于所述第一编码结果得到的。The second sending module 36 is configured to send the first encoding result to the first communication device, and the first channel information is based on the first encoding result by the first communication device through the first decoder owned.
在一个可选的实施例中,In an alternative embodiment,
所述第一无线信道模型包括第一信道估计模型;The first wireless channel model includes a first channel estimation model;
所述第一信道信息包括第一信道估计结果;The first channel information includes a first channel estimation result;
所述第二信道信息包括参考信号。The second channel information includes a reference signal.
在一个可选的实施例中,In an alternative embodiment,
所述第一无线信道模型包括第一CSI自编码模型,所述第二无线信道模型包括第二CSI自编码模型;The first wireless channel model includes a first CSI self-encoding model, and the second wireless channel model includes a second CSI self-encoding model;
所述第一信道信息包括第一CSI恢复结果。The first channel information includes a first CSI recovery result.
在一个可选的实施例中,所述第二更新模块32,用于在基于第一概率值确定第一无线信道模型满足更新条件的情况下,更新所述第一无线信道模型,所述第一概率值是通过所述第一判别器基于所述第一信道信息得到的。In an optional embodiment, the second update module 32 is configured to update the first wireless channel model when it is determined based on the first probability value that the first wireless channel model satisfies the update condition, and the first A probability value is obtained by the first discriminator based on the first channel information.
在一个可选的实施例中,所述第一信道信息包括多个信道信息,所述第一概率值包括与所述多个信道信息一一对应的多个概率值;In an optional embodiment, the first channel information includes a plurality of channel information, and the first probability value includes a plurality of probability values corresponding to the plurality of channel information one-to-one;
所述更新条件包括如下条件中的至少一种:The update conditions include at least one of the following conditions:
所述第一概率值中概率值低于第一阈值的占比高于预设值;The proportion of probability values lower than the first threshold among the first probability values is higher than a preset value;
所述第一概率值中连续的x个概率值低于第二阈值,x为正整数;x consecutive probability values in the first probability value are lower than the second threshold, and x is a positive integer;
根据所述第一概率值得到的概率值分布情况满足第一条件;The probability value distribution obtained according to the first probability value satisfies the first condition;
基于所述第一概率值计算得到的评价数值达到第三阈值。The evaluation value calculated based on the first probability value reaches a third threshold.
在一个可选的实施例中,所述第一判别器包括至少两个子判别器,所述第一概率值包括至少两个子概率值。In an optional embodiment, the first discriminator includes at least two sub-discriminators, and the first probability value includes at least two sub-probability values.
在一个可选的实施例中,所述第二更新模块32,用于在基于所述至少两个子概率值中的至少一个子概率值确定所述第一无线信道模型满足更新条件的情况下,更新所述第一无线信道模型。In an optional embodiment, the second update module 32 is configured to determine that the first wireless channel model satisfies an update condition based on at least one sub-probability value of the at least two sub-probability values, Updating the first wireless channel model.
图23示出了本申请一个示例性实施例提供的通信设备(终端或网络设备)的结构示意图,该通信设备包括:处理器1001、接收器1002、发射器1003、存储器1004。FIG. 23 shows a schematic structural diagram of a communication device (terminal or network device) provided by an exemplary embodiment of the present application. The communication device includes: a processor 1001 , a receiver 1002 , a transmitter 1003 , and a memory 1004 .
处理器1001包括一个或者一个以上处理核心,处理器1001通过运行软件程序以及模块,从而执行各种功能应用以及信息处理。The processor 1001 includes one or more processing cores, and the processor 1001 executes various functional applications and information processing by running software programs and modules.
接收器1002和发射器1003可以实现为一个通信组件,该通信组件可以是一块通信芯片。The receiver 1002 and the transmitter 1003 can be realized as a communication component, and the communication component can be a communication chip.
存储器1004与处理器1001相连。The memory 1004 is connected to the processor 1001 .
存储器1004可用于存储至少一个指令,处理器1001用于执行该至少一个指令,以实现上述方法实施例中的各个步骤。The memory 1004 may be used to store at least one instruction, and the processor 1001 is used to execute the at least one instruction, so as to implement various steps in the foregoing method embodiments.
此外,存储器1004可以由任何类型的易失性或非易失性存储设备或者它们的组合实现,易失性或非易失性存储设备包括但不限于:磁盘或光盘,电可擦除可编程只读存储器(Electrically-Erasable Programmable Read Only Memory,EEPROM),可擦除可编程只读存储器(Erasable Programmable Read Only Memory,EPROM),静态随时存取存储器(Static Random Access Memory,SRAM),只读存储器(Read-Only Memory,ROM),磁存储器,快闪存储器,可编程只读存储器(Programmable Read-Only Memory,PROM)。In addition, the memory 1004 can be realized by any type of volatile or non-volatile storage device or their combination, volatile or non-volatile storage devices include but not limited to: magnetic disk or optical disk, electrically erasable and programmable Electrically-Erasable Programmable Read Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Static Random Access Memory (SRAM), Read-Only Memory (Read-Only Memory, ROM), magnetic memory, flash memory, programmable read-only memory (Programmable Read-Only Memory, PROM).
在示例性实施例中,还提供了一种计算机可读存储介质,所述计算机可读存储介质中存储有至少一条指令、至少一段程序、代码集或指令集,所述至少一条指令、所述至少一段程序、所述代码集或指令集由所述处理器加载并执行以实现上述各个方法实施例提供的由终端或网络设备执行的无线信道模型的更新方法。In an exemplary embodiment, a computer-readable storage medium is also provided, the computer-readable storage medium stores at least one instruction, at least one program, a code set or an instruction set, the at least one instruction, the At least one program, the code set or the instruction set is loaded and executed by the processor to implement the method for updating the wireless channel model performed by the terminal or network device provided in the above method embodiments.
在示例性实施例中,还提供了一种计算机程序产品或计算机程序,该计算机程序产品或计算机程序包括计算机指令,该计算机指令存储在计算机可读存储介质中,通信设备的处理器从计算机可读存储介质读取该计算机指令,处理器执行该计算机指令,使得该通信设备执行上述方面所述的由终端或网络设备执行的无线信道模型的更新方法。In an exemplary embodiment, there is also provided a computer program product or computer program, the computer program product or computer program comprising computer instructions, the computer instructions are stored in a computer-readable storage medium, the processor of the communication device can read from the computer The computer instruction is read by reading the storage medium, and the processor executes the computer instruction, so that the communication device executes the method for updating the wireless channel model performed by the terminal or network device described in the above aspect.
以上所述仅为本申请的可选实施例,并不用以限制本申请,凡在本申请的精神和原则之内,所作的任何修改、等同替换、改进等,均应包含在本申请的保护范围之内。The above are only optional embodiments of the application, and are not intended to limit the application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the application shall be included in the protection of the application. within range.
Claims (67)
- 一种无线信道模型的更新方法,其特征在于,所述方法由第一通信设备执行,所述方法包括:A method for updating a wireless channel model, characterized in that the method is executed by a first communication device, and the method includes:通过第一无线信道模型得到第一信道信息,所述第一无线信道模型为第一生成对抗网络中的生成器;Obtaining first channel information through a first wireless channel model, where the first wireless channel model is a generator in a first generation confrontation network;在通过第一判别器基于所述第一信道信息确定所述第一无线信道模型满足更新条件的情况下,触发更新所述第一无线信道模型,所述第一判别器为所述第一生成对抗网络中的判别器。When it is determined by the first discriminator based on the first channel information that the first wireless channel model satisfies the update condition, an update of the first wireless channel model is triggered, and the first discriminator generates for the first Discriminator in adversarial networks.
- 根据权利要求1所述的方法,其特征在于,所述触发更新所述第一无线信道模型,包括:The method according to claim 1, wherein the triggering to update the first wireless channel model comprises:向第二通信设备发送第一更新指示,所述第一更新指示用于触发所述第二通信设备更新所述第一无线信道模型得到第二无线信道模型;sending a first update instruction to a second communication device, where the first update instruction is used to trigger the second communication device to update the first wireless channel model to obtain a second wireless channel model;接收所述第二通信设备发送的所述第二无线信道模型。Receive the second wireless channel model sent by the second communication device.
- 根据权利要求2所述的方法,其特征在于,所述通过第一无线信道模型得到第一信道信息,包括:The method according to claim 2, wherein said obtaining the first channel information through the first wireless channel model comprises:通过所述第一无线信道模型基于第二信道信息得到所述第一信道信息;obtaining the first channel information based on the second channel information through the first wireless channel model;所述方法还包括:The method also includes:将所述第二信道信息作为训练样本实时训练更新所述第一判别器得到第二判别器;Using the second channel information as a training sample to train and update the first discriminator in real time to obtain a second discriminator;其中,所述第一更新指示包括所述第二判别器;所述第二判别器用于与所述第一无线信道模型组成第二生成对抗网络训练更新所述第一无线信道模型。Wherein, the first update instruction includes the second discriminator; the second discriminator is used to form a second generative adversarial network with the first wireless channel model to train and update the first wireless channel model.
- 根据权利要求1所述的方法,其特征在于,所述触发更新所述第一无线信道模型,包括:The method according to claim 1, wherein the triggering to update the first wireless channel model comprises:更新所述第一无线信道模型得到第二无线信道模型。Updating the first wireless channel model to obtain a second wireless channel model.
- 根据权利要求4所述的方法,其特征在于,所述通过第一无线信道模型得到第一信道信息,包括:The method according to claim 4, wherein said obtaining the first channel information through the first wireless channel model comprises:通过所述第一无线信道模型基于第二信道信息得到所述第一信道信息;obtaining the first channel information based on the second channel information through the first wireless channel model;所述方法还包括:The method also includes:将所述第二信道信息作为训练样本实时训练更新所述第一判别器得到第二判别器;Using the second channel information as a training sample to train and update the first discriminator in real time to obtain a second discriminator;所述更新所述第一无线信道模型得到第二无线信道模型,包括:The updating the first wireless channel model to obtain a second wireless channel model includes:基于所述第二判别器和所述第一无线信道模型组成的第二生成对抗网络,训练更新所述第一无线信道模型得到所述第二无线信道模型。Based on the second generative adversarial network composed of the second discriminator and the first wireless channel model, train and update the first wireless channel model to obtain the second wireless channel model.
- 根据权利要求4所述的方法,其特征在于,所述更新所述第一无线信道模型得到第二无线信道模型,包括:The method according to claim 4, wherein said updating said first wireless channel model to obtain a second wireless channel model comprises:基于所述第一判别器和所述第一无线信道模型组成的所述第一生成对抗网络,训练更新所述第一判别器和所述第一无线信道模型得到第二判别器和所述第二无线信道模型。Based on the first generative confrontation network composed of the first discriminator and the first wireless channel model, train and update the first discriminator and the first wireless channel model to obtain a second discriminator and the first wireless channel model Two wireless channel models.
- 根据权利要求1至6任一所述的方法,其特征在于,所述方法还包括:The method according to any one of claims 1 to 6, wherein the method further comprises:接收第二通信设备发送的第二信道信息;receiving second channel information sent by the second communication device;所述通过第一无线信道模型得到第一信道信息,包括:The obtaining the first channel information through the first wireless channel model includes:通过所述第一无线信道模型基于所述第二信道信息得到所述第一信道信息。The first channel information is obtained based on the second channel information through the first wireless channel model.
- 根据权利要求2所述的方法,其特征在于,所述第一无线信道模型包括第一编码器和第一解码器,所述第一解码器部署在所述第一通信设备侧,所述第一编码器部署在所述第二通信设备侧;所述第二无线信道模型包括第二编码器和第二解码器;The method according to claim 2, wherein the first wireless channel model includes a first encoder and a first decoder, the first decoder is deployed on the side of the first communication device, and the first An encoder is deployed on the side of the second communication device; the second wireless channel model includes a second encoder and a second decoder;所述接收所述第二通信设备发送的所述第二无线信道模型,包括:The receiving the second wireless channel model sent by the second communication device includes:接收所述第二通信设备发送的所述第二解码器和第二判别器。Receive the second decoder and the second discriminator sent by the second communication device.
- 根据权利要求8所述的方法,其特征在于,所述方法还包括:The method according to claim 8, characterized in that the method further comprises:接收所述第二通信设备发送的第一编码结果,所述第一编码结果是通过所述第一编码器基于第二信道信息得到的;receiving a first encoding result sent by the second communication device, where the first encoding result is obtained by the first encoder based on second channel information;所述通过第一无线信道模型得到第一信道信息,包括:The obtaining the first channel information through the first wireless channel model includes:通过所述第一解码器基于所述第一编码结果得到所述第一信道信息。The first channel information is obtained based on the first encoding result by the first decoder.
- 根据权利要求5或6所述的方法,其特征在于,所述第一无线信道模型包括第一编码器和第一解码器,所述第一编码器部署在所述第一通信设备侧,所述第一解码器部署在第二通信设备侧,所述第一通信设备侧存储有所述第一解码器;所述第二无线信道模型包括第二编码器和第二解码器;The method according to claim 5 or 6, wherein the first wireless channel model includes a first encoder and a first decoder, and the first encoder is deployed on the side of the first communication device, so The first decoder is deployed on the side of the second communication device, and the first decoder is stored on the side of the first communication device; the second wireless channel model includes a second encoder and a second decoder;所述方法还包括:The method also includes:向所述第二通信设备发送所述第二解码器和所述第二判别器。sending the second decoder and the second discriminator to the second communications device.
- 根据权利要求7所述的方法,其特征在于,所述第一通信设备包括终端设备,所述第二通信设备包括网络设备;或,所述第一通信设备包括网络设备,所述第一通信设备包括终端设备;The method according to claim 7, wherein the first communication device includes a terminal device, and the second communication device includes a network device; or, the first communication device includes a network device, and the first communication device Equipment includes terminal equipment;所述第一无线信道模型包括第一信道估计模型;The first wireless channel model includes a first channel estimation model;所述第一信道信息包括第一信道估计结果;The first channel information includes a first channel estimation result;所述第二信道信息包括参考信号。The second channel information includes a reference signal.
- 根据权利要求8或9所述的方法,其特征在于,所述第一通信设备包括网络设备,所述第二通信设备包括终端设备;The method according to claim 8 or 9, wherein the first communication device comprises a network device, and the second communication device comprises a terminal device;所述第一无线信道模型包括第一信道状态信息CSI自编码模型,所述第二无线信道模型包括第二CSI自编码模型;The first wireless channel model includes a first channel state information CSI self-encoding model, and the second wireless channel model includes a second CSI self-encoding model;所述第一信道信息包括第一CSI恢复结果。The first channel information includes a first CSI restoration result.
- 根据权利要求10所述的方法,其特征在于,所述第一通信设备包括终端设备,所述第二通信设备包括网络设备;The method according to claim 10, wherein the first communication device comprises a terminal device, and the second communication device comprises a network device;所述第一无线信道模型包括第一CSI预测自编码模型,所述第二无线信道模型包括第二CSI预测自编码模型;The first wireless channel model includes a first CSI predictive self-encoding model, and the second wireless channel model includes a second CSI predictive self-encoding model;所述第一信道信息包括第一CSI预测结果。The first channel information includes a first CSI prediction result.
- 根据权利要求1至13任一所述的方法,其特征在于,所述在通过第一判别器基于所述第一信道信息确定所述第一无线信道模型满足更新条件的情况下,触发更新所述第一无线信道模型,包括:The method according to any one of claims 1 to 13, wherein, when the first discriminator determines that the first wireless channel model satisfies the update condition based on the first channel information, triggering the update of the The first wireless channel model includes:通过所述第一判别器评价所述第一信道信息得到第一概率值;evaluating the first channel information by the first discriminator to obtain a first probability value;在基于所述第一概率值确定所述第一无线信道模型满足更新条件的情况下,触发更新所述第一无线信道模型。In a case where it is determined based on the first probability value that the first wireless channel model satisfies an update condition, an update of the first wireless channel model is triggered.
- 根据权利要求14所述的方法,其特征在于,所述第一信道信息包括多个信道信息,所述第一概率值包括与所述多个信道信息一一对应的多个概率值;The method according to claim 14, wherein the first channel information includes a plurality of channel information, and the first probability value includes a plurality of probability values corresponding one to one to the plurality of channel information;所述更新条件包括如下条件中的至少一种:The update conditions include at least one of the following conditions:所述第一概率值中概率值低于第一阈值的占比高于预设值;The proportion of probability values lower than the first threshold among the first probability values is higher than a preset value;所述第一概率值中连续的x个概率值低于第二阈值,x为正整数;x consecutive probability values in the first probability value are lower than the second threshold, and x is a positive integer;根据所述第一概率值得到的概率值分布情况满足第一条件;The probability value distribution obtained according to the first probability value satisfies the first condition;基于所述第一概率值计算得到的评价数值达到第三阈值。The evaluation value calculated based on the first probability value reaches a third threshold.
- 根据权利要求14所述的方法,其特征在于,所述第一判别器包括至少两个子判别器,所述第一概率值包括至少两个子概率值。The method according to claim 14, wherein the first discriminator comprises at least two sub-discriminators, and the first probability value comprises at least two sub-probability values.
- 根据权利要求16所述的方法,其特征在于,所述在基于所述第一概率值确定所述第一无线信道模型满足更新条件的情况下,触发更新所述第一无线信道模型,包括:The method according to claim 16, wherein the triggering to update the first wireless channel model in a case where it is determined based on the first probability value that the first wireless channel model satisfies an update condition comprises:在基于所述至少两个子概率值中的至少一个子概率值确定所述第一无线信道模型满足更新条件的情况下,触发更新所述第一无线信道模型。In a case where it is determined based on at least one sub-probability value of the at least two sub-probability values that the first radio channel model satisfies an update condition, triggering an update of the first radio channel model.
- 一种无线信道模型的更新方法,其特征在于,所述方法由第二通信设备执行,所述方法包括:A method for updating a wireless channel model, characterized in that the method is performed by a second communication device, and the method includes:在通过第一判别器基于第一信道信息确定第一无线信道模型满足更新条件的情况下,更新所述第一无线信道模型,所述第一信道信息是所述第一无线信道模型输出的,所述第一无线信道模型为第一生成对抗网络中的生成器,所述第一判别器为所述第一生成对抗网络中的判别器。updating the first wireless channel model when it is determined by the first discriminator based on the first channel information that the first wireless channel model satisfies the update condition, the first channel information is output by the first wireless channel model, The first wireless channel model is a generator in the first generative adversarial network, and the first discriminator is a discriminator in the first generative adversarial network.
- 根据权利要求18所述的方法,其特征在于,所述在通过第一判别器基于第一信道信息确定第一无线信道模型满足更新条件的情况下,更新所述第一无线信道模型,包括:The method according to claim 18, wherein the updating of the first wireless channel model includes:接收第一通信设备发送的第一更新指示,所述第一更新指示是所述第一通信设备在通过所述第一判别器基于所述第一信道信息确定所述第一无线信道模型满足更新条件的情况下发送的;Receiving a first update indication sent by the first communication device, where the first update indication is that the first communication device determines that the first wireless channel model satisfies the update requirement based on the first channel information through the first discriminator. sent under condition;更新所述第一无线信道模型得到第二无线信道模型。Updating the first wireless channel model to obtain a second wireless channel model.
- 根据权利要求19所述的方法,其特征在于,所述方法还包括:The method according to claim 19, further comprising:向所述第一通信设备发送所述第二无线信道模型。The second wireless channel model is sent to the first communications device.
- 根据权利要求19所述的方法,其特征在于,所述第一更新指示包括第二判别器,所述第二判别器是使用第二信道信息实时训练更新所述第一判别器得到的,所述第二信道信息是所述第一无线信道模型生成所述第一信道信息时的输入信息;The method according to claim 19, wherein the first update instruction includes a second discriminator, and the second discriminator is obtained by using the second channel information to train and update the first discriminator in real time, so The second channel information is input information when the first wireless channel model generates the first channel information;所述更新所述第一无线信道模型得到第二无线信道模型,包括;The updating the first wireless channel model to obtain a second wireless channel model includes;在所述第二判别器与所述第一无线信道模型组成的第二生成对抗网络中,训练更新所述第一无线信道模型得到所述第二无线信道模型。In the second generative adversarial network composed of the second discriminator and the first wireless channel model, train and update the first wireless channel model to obtain the second wireless channel model.
- 根据权利要求18至21任一所述的方法,其特征在于,所述方法还包括:The method according to any one of claims 18 to 21, further comprising:向所述第一通信设备发送第二信道信息,所述第二信道信息是所述第一无线信道模型生成所述第一信道信息时的输入信息。Sending second channel information to the first communication device, where the second channel information is input information when the first wireless channel model generates the first channel information.
- 根据权利要求20所述的方法,其特征在于,所述第一无线信道模型包括第一编码器和第一解码器,所述第一解码器部署在所述第一通信设备侧,所述第一编码器部署在所述第二通信设备侧;所述第二无线信道模型包括第二编码器和第二解码器;The method according to claim 20, wherein the first wireless channel model includes a first encoder and a first decoder, the first decoder is deployed on the side of the first communication device, and the first An encoder is deployed on the side of the second communication device; the second wireless channel model includes a second encoder and a second decoder;所述向所述第一通信设备发送所述第二无线信道模型,包括:The sending the second wireless channel model to the first communication device includes:向所述第一通信设备发送所述第二解码器和第二判别器。The second decoder and second discriminator are sent to the first communications device.
- 根据权利要求23所述的方法,其特征在于,所述方法还包括:The method according to claim 23, further comprising:通过所述第一编码器基于第二信道信息得到第一编码结果,所述第二信道信息是所述第一无线信道模型生成所述第一信道信息时的输入信息;Obtaining a first encoding result by the first encoder based on second channel information, where the second channel information is input information when the first wireless channel model generates the first channel information;向所述第一通信设备发送所述第一编码结果,所述第一信道信息是所述第一通信设备通过所述第一解码器基于所述第一编码结果得到的。sending the first encoding result to the first communication device, where the first channel information is obtained by the first communication device through the first decoder based on the first encoding result.
- 根据权利要求22所述的方法,其特征在于,所述第一通信设备包括终端设备,所述第二通信设备包括网络设备;或,所述第一通信设备包括网络设备,所述第一通信设备包括终端设备;The method according to claim 22, wherein the first communication device includes a terminal device, and the second communication device includes a network device; or, the first communication device includes a network device, and the first communication device Equipment includes terminal equipment;所述第一无线信道模型包括第一信道估计模型;The first wireless channel model includes a first channel estimation model;所述第一信道信息包括第一信道估计结果;The first channel information includes a first channel estimation result;所述第二信道信息包括参考信号。The second channel information includes a reference signal.
- 根据权利要求23或24所述的方法,其特征在于,所述第一通信设备包括网络设备,所述第二通信设备包括终端设备;The method according to claim 23 or 24, wherein the first communication device comprises a network device, and the second communication device comprises a terminal device;所述第一无线信道模型包括第一信道状态信息CSI自编码模型,所述第二无线信道模型包括第二CSI自编码模型;The first wireless channel model includes a first channel state information CSI self-encoding model, and the second wireless channel model includes a second CSI self-encoding model;所述第一信道信息包括第一CSI恢复结果。The first channel information includes a first CSI recovery result.
- 根据权利要求18至26任一所述的方法,其特征在于,所述在通过第一判别器基于所述第一信道信息确定所述第一无线信道模型满足更新条件的情况下,更新所述第一无线信道模型,包括:The method according to any one of claims 18 to 26, wherein when the first discriminator determines that the first wireless channel model satisfies the update condition based on the first channel information, updating the The first wireless channel model, including:在基于第一概率值确定所述第一无线信道模型满足更新条件的情况下,更新所述第一无线信道模型,所述第一概率值是通过所述第一判别器基于所述第一信道信息得到的。Update the first wireless channel model when it is determined that the first wireless channel model satisfies the update condition based on the first probability value, the first probability value is based on the first channel through the first discriminator information obtained.
- 根据权利要求27所述的方法,其特征在于,所述第一信道信息包括多个信道信息,所述第一概率值包括与所述多个信道信息一一对应的多个概率值;The method according to claim 27, wherein the first channel information includes a plurality of channel information, and the first probability value includes a plurality of probability values corresponding to the plurality of channel information one-to-one;所述更新条件包括如下条件中的至少一种:The update conditions include at least one of the following conditions:所述第一概率值中概率值低于第一阈值的占比高于预设值;The proportion of probability values lower than the first threshold among the first probability values is higher than a preset value;所述第一概率值中连续的x个概率值低于第二阈值,x为正整数;x consecutive probability values in the first probability value are lower than the second threshold, and x is a positive integer;根据所述第一概率值得到的概率值分布情况满足第一条件;The probability value distribution obtained according to the first probability value satisfies the first condition;基于所述第一概率值计算得到的评价数值达到第三阈值。The evaluation value calculated based on the first probability value reaches a third threshold.
- 根据权利要求27所述的方法,其特征在于,所述第一判别器包括至少两个子判别器,所述第一概率值包括至少两个子概率值。The method according to claim 27, wherein the first discriminator comprises at least two sub-discriminators, and the first probability value comprises at least two sub-probability values.
- 根据权利要求29所述的方法,其特征在于,所述在基于第一概率值确定所述第一无线信道模型满足更新条件的情况下,触发更新所述第一无线信道模型,包括:The method according to claim 29, wherein the triggering to update the first wireless channel model in the case of determining that the first wireless channel model satisfies an update condition based on the first probability value comprises:在基于所述至少两个子概率值中的至少一个子概率值确定所述第一无线信道模型满足更新条件的情况下,触发更新所述第一无线信道模型。In a case where it is determined based on at least one sub-probability value of the at least two sub-probability values that the first radio channel model satisfies an update condition, triggering an update of the first radio channel model.
- 一种第一通信装置,其特征在于,所述装置包括:A first communication device, characterized in that the device includes:第一模型模块,用于通过第一无线信道模型得到第一信道信息,所述第一无线信道模型为第一生成对抗网络中的生成器;The first model module is used to obtain the first channel information through the first wireless channel model, and the first wireless channel model is a generator in the first generative confrontation network;第一更新模块,用于在通过第一判别器基于所述第一信道信息确定所述第一无线信道模型满足更新条件的情况下,触发更新所述第一无线信道模型,所述第一判别器为所述第一生成对抗网络中的判别器。A first updating module, configured to trigger an update of the first wireless channel model when the first discriminator determines based on the first channel information that the first wireless channel model satisfies an update condition, the first discriminator The discriminator is the discriminator in the first generative confrontation network.
- 根据权利要求31所述的装置,其特征在于,所述第一更新模块包括:The device according to claim 31, wherein the first update module comprises:第一发送子模块,用于向第二通信装置发送第一更新指示,所述第一更新指示用于触发所述第二通信装置更新所述第一无线信道模型得到第二无线信道模型;A first sending submodule, configured to send a first update instruction to a second communication device, where the first update instruction is used to trigger the second communication device to update the first wireless channel model to obtain a second wireless channel model;第一接收子模块,用于接收所述第二通信装置发送的所述第二无线信道模型。The first receiving submodule is configured to receive the second wireless channel model sent by the second communication device.
- 根据权利要求32所述的装置,其特征在于,The device according to claim 32, characterized in that,所述第一模型模块,用于通过所述第一无线信道模型基于第二信道信息得到所述第一信道信息;The first model module is configured to obtain the first channel information based on the second channel information through the first wireless channel model;所述装置还包括:The device also includes:第一实时训练模块,用于将所述第二信道信息作为训练样本实时训练更新所述第一判别器得到第二判别器;The first real-time training module is used to use the second channel information as a training sample to train and update the first discriminator in real time to obtain a second discriminator;其中,所述第一更新指示包括所述第二判别器;所述第二判别器用于与所述第一无线信道模型组成第二生成对抗网络训练更新所述第一无线信道模型。Wherein, the first update instruction includes the second discriminator; the second discriminator is used to form a second generative adversarial network with the first wireless channel model to train and update the first wireless channel model.
- 根据权利要求31所述的装置,其特征在于,所述第一更新模块,用于更新所述第一无线信道模型得到第二无线信道模型。The device according to claim 31, wherein the first updating module is configured to update the first wireless channel model to obtain a second wireless channel model.
- 根据权利要求34所述的装置,其特征在于,所述第一模型模块,用于通过所述第一无线信道模型基于第二信道信息得到所述第一信道信息;The device according to claim 34, wherein the first model module is configured to obtain the first channel information based on the second channel information through the first wireless channel model;所述装置还包括:The device also includes:第一实时训练模块,用于将所述第二信道信息作为训练样本实时训练更新所述第一判别器得到第二判别器;The first real-time training module is used to use the second channel information as a training sample to train and update the first discriminator in real time to obtain a second discriminator;所述第一更新模块,用于基于所述第二判别器和所述第一无线信道模型组成的第二生成对抗网络,训练更新所述第一无线信道模型得到所述第二无线信道模型。The first update module is configured to train and update the first wireless channel model to obtain the second wireless channel model based on the second generative confrontation network composed of the second discriminator and the first wireless channel model.
- 根据权利要求34所述的装置,其特征在于,所述第一更新模块,用于基于所述第一判别器和所述第一无线信道模型组成的所述第一生成对抗网络,训练更新所述第一判别器和所述第一无线信道模型得到第二判别器和所述第二无线信道模型。The device according to claim 34, wherein the first update module is configured to train and update the first generative adversarial network based on the first discriminator and the first wireless channel model. The first discriminator and the first wireless channel model are used to obtain a second discriminator and the second wireless channel model.
- 根据权利要求31至36任一所述的装置,其特征在于,所述装置还包括:The device according to any one of claims 31 to 36, wherein the device further comprises:第一接收模块,用于接收第二通信装置发送的第二信道信息;The first receiving module is configured to receive the second channel information sent by the second communication device;所述第一模型模块,用于通过所述第一无线信道模型基于所述第二信道信息得到所述第一信道信息。The first model module is configured to obtain the first channel information based on the second channel information through the first wireless channel model.
- 根据权利要求32所述的装置,其特征在于,所述第一无线信道模型包括第一编码器和第一解码器,所述第一解码器部署在所述第一通信装置侧,所述第一编码器部署在所述第二通信装置侧;所述第二无线信道模型包括第二编码器和第二解码器;The device according to claim 32, wherein the first wireless channel model includes a first encoder and a first decoder, the first decoder is deployed on the side of the first communication device, and the first An encoder is deployed on the side of the second communication device; the second wireless channel model includes a second encoder and a second decoder;所述第一接收子模块,用于接收所述第二通信装置发送的所述第二解码器和第二判别器。The first receiving submodule is configured to receive the second decoder and the second discriminator sent by the second communication device.
- 根据权利要求38所述的装置,其特征在于,所述装置还包括:The device according to claim 38, further comprising:第一接收模块,用于接收所述第二通信装置发送的第一编码结果,所述第一编码结果是通过所述第一编码器基于第二信道信息得到的;A first receiving module, configured to receive a first encoding result sent by the second communication device, where the first encoding result is obtained by the first encoder based on the second channel information;所述第一模型模块,用于通过所述第一解码器基于所述第一编码结果得到所述第一信道信息。The first model module is configured to use the first decoder to obtain the first channel information based on the first encoding result.
- 根据权利要求35或36所述的装置,其特征在于,所述第一无线信道模型包括第一编码器和第一解码器,所述第一编码器部署在所述第一通信装置侧,所述第一解码器部署在第二通信装置侧,所述第一通信装置侧存储有所述第一解码器;所述第二无线信道模型包括第二编码器和第二解码器;The device according to claim 35 or 36, wherein the first wireless channel model includes a first encoder and a first decoder, and the first encoder is deployed on the side of the first communication device, so The first decoder is deployed on the side of the second communication device, and the first decoder is stored on the side of the first communication device; the second wireless channel model includes a second encoder and a second decoder;所述第一更新模块包括:The first update module includes:第一发送子模块,用于向所述第二通信装置发送所述第二解码器和所述第二判别器。The first sending submodule is configured to send the second decoder and the second discriminator to the second communication device.
- 根据权利要求37所述的装置,其特征在于,Apparatus according to claim 37, characterized in that所述第一无线信道模型包括第一信道估计模型;The first wireless channel model includes a first channel estimation model;所述第一信道信息包括第一信道估计结果;The first channel information includes a first channel estimation result;所述第二信道信息包括参考信号。The second channel information includes a reference signal.
- 根据权利要求38或39所述的装置,其特征在于,Apparatus according to claim 38 or 39, characterized in that所述第一无线信道模型包括第一信道状态信息CSI自编码模型,所述第二无线信道模型包括第二CSI自编码模型;The first wireless channel model includes a first channel state information CSI self-encoding model, and the second wireless channel model includes a second CSI self-encoding model;所述第一信道信息包括第一CSI恢复结果。The first channel information includes a first CSI recovery result.
- 根据权利要求40所述的装置,其特征在于,The device according to claim 40, characterized in that,所述第一无线信道模型包括第一CSI预测自编码模型,所述第二无线信道模型包括第二CSI预测自编码模型;The first wireless channel model includes a first CSI predictive self-encoding model, and the second wireless channel model includes a second CSI predictive self-encoding model;所述第一信道信息包括第一CSI预测结果。The first channel information includes a first CSI prediction result.
- 根据权利要求31至43任一所述的装置,其特征在于,所述更新模块,包括;The device according to any one of claims 31 to 43, wherein the update module includes;第一评价子模块,用于通过所述第一判别器评价所述第一信道信息得到第一概率值;A first evaluation submodule, configured to evaluate the first channel information through the first discriminator to obtain a first probability value;第一更新子模块,用于在基于所述第一概率值确定所述第一无线信道模型满足更新条件的情况下,触发更新所述第一无线信道模型。The first update submodule is configured to trigger an update of the first wireless channel model when it is determined based on the first probability value that the first wireless channel model satisfies an update condition.
- 根据权利要求44所述的装置,其特征在于,所述第一信道信息包括多个信道信息,所述第一概率值包括与所述多个信道信息一一对应的多个概率值;The device according to claim 44, wherein the first channel information includes a plurality of channel information, and the first probability value includes a plurality of probability values corresponding to the plurality of channel information one-to-one;所述更新条件包括如下条件中的至少一种:The update conditions include at least one of the following conditions:所述第一概率值中概率值低于第一阈值的占比高于预设值;The proportion of probability values lower than the first threshold among the first probability values is higher than a preset value;所述第一概率值中连续的x个概率值低于第二阈值,x为正整数;x consecutive probability values in the first probability value are lower than the second threshold, and x is a positive integer;根据所述第一概率值得到的概率值分布情况满足第一条件;The probability value distribution obtained according to the first probability value satisfies the first condition;基于所述第一概率值计算得到的评价数值达到第三阈值。The evaluation value calculated based on the first probability value reaches a third threshold.
- 根据权利要求44所述的装置,其特征在于,所述第一判别器包括至少两个子判别器,所述第一概率值包括至少两个子概率值。The apparatus according to claim 44, wherein the first discriminator comprises at least two sub-discriminators, and the first probability value comprises at least two sub-probability values.
- 根据权利要求46所述的装置,其特征在于,所述第一更新子模块,用于在基于所述至少两个子概 率值中的至少一个子概率值确定所述第一无线信道模型满足更新条件的情况下,触发更新所述第一无线信道模型。The device according to claim 46, wherein the first update submodule is configured to determine that the first wireless channel model satisfies the update condition based on at least one sub-probability value of the at least two sub-probability values In the case of , trigger to update the first wireless channel model.
- 一种第二通信装置,其特征在于,所述装置包括:A second communication device, characterized in that the device includes:第二更新模块,用于在通过第一判别器基于第一信道信息确定第一无线信道模型满足更新条件的情况下,更新所述第一无线信道模型,所述第一信道信息是所述第一无线信道模型输出的,所述第一无线信道模型为第一生成对抗网络中的生成器,所述第一判别器为所述第一生成对抗网络中的判别器。The second update module is configured to update the first wireless channel model when the first discriminator determines that the first wireless channel model satisfies the update condition based on the first channel information, the first channel information being the first channel information A wireless channel model is output, the first wireless channel model is a generator in the first generative adversarial network, and the first discriminator is a discriminator in the first generative adversarial network.
- 根据权利要求48所述的装置,其特征在于,所述第二更新模块,包括:The device according to claim 48, wherein the second updating module comprises:第二接收子模块,用于接收第一通信装置发送的第一更新指示,所述第一更新指示是所述第一通信装置在通过所述第一判别器基于所述第一信道信息确定所述第一无线信道模型满足更新条件的情况下发送的;The second receiving submodule is configured to receive a first update instruction sent by the first communication device, where the first update instruction is determined by the first communication device based on the first channel information through the first discriminator sent when the above-mentioned first wireless channel model satisfies the update condition;第二更新子模块,用于更新所述第一无线信道模型得到第二无线信道模型。The second updating submodule is configured to update the first wireless channel model to obtain a second wireless channel model.
- 根据权利要求49所述的装置,其特征在于,所述第二更新模块,包括:The device according to claim 49, wherein the second update module includes:第二发送子模块,用于向所述第一通信装置发送所述第二无线信道模型。The second sending submodule is configured to send the second wireless channel model to the first communication device.
- 根据权利要求49所述的装置,其特征在于,所述第一更新指示包括第二判别器,所述第二判别器是使用第二信道信息实时训练更新所述第一判别器得到的,所述第二信道信息是所述第一无线信道模型生成所述第一信道信息时的输入信息;The device according to claim 49, wherein the first update instruction includes a second discriminator, and the second discriminator is obtained by using the second channel information to train and update the first discriminator in real time, so The second channel information is input information when the first wireless channel model generates the first channel information;所述第二更新子模块,用于在所述第二判别器与所述第一无线信道模型组成的第二生成对抗网络中,训练更新所述第一无线信道模型得到所述第二无线信道模型。The second updating submodule is configured to train and update the first wireless channel model to obtain the second wireless channel in the second generative confrontation network composed of the second discriminator and the first wireless channel model Model.
- 根据权利要求48至51任一所述的装置,其特征在于,所述装置还包括:The device according to any one of claims 48 to 51, wherein the device further comprises:第二发送模块,用于向所述第一通信装置发送第二信道信息,所述第二信道信息是所述第一无线信道模型生成所述第一信道信息时的输入信息。A second sending module, configured to send second channel information to the first communication device, where the second channel information is input information when the first wireless channel model generates the first channel information.
- 根据权利要求50所述的装置,其特征在于,所述第一无线信道模型包括第一编码器和第一解码器,所述第一解码器部署在所述第一通信装置侧,所述第一编码器部署在所述第二通信装置侧;所述第二无线信道模型包括第二编码器和第二解码器;The device according to claim 50, wherein the first wireless channel model includes a first encoder and a first decoder, the first decoder is deployed on the side of the first communication device, and the first An encoder is deployed on the side of the second communication device; the second wireless channel model includes a second encoder and a second decoder;所述第二发送子模块,用于向所述第一通信装置发送所述第二解码器和第二判别器。The second sending submodule is configured to send the second decoder and the second discriminator to the first communication device.
- 根据权利要求53所述的装置,其特征在于,所述装置还包括:The device according to claim 53, further comprising:第二模型模块,用于通过所述第一编码器基于第二信道信息得到第一编码结果,所述第二信道信息是所述第一无线信道模型生成所述第一信道信息时的输入信息;A second model module, configured to use the first encoder to obtain a first encoding result based on second channel information, where the second channel information is input information when the first wireless channel model generates the first channel information ;第二发送模块,用于向所述第一通信装置发送所述第一编码结果,所述第一信道信息是所述第一通信装置通过所述第一解码器基于所述第一编码结果得到的。A second sending module, configured to send the first encoding result to the first communication device, where the first channel information is obtained by the first communication device through the first decoder based on the first encoding result of.
- 根据权利要求52所述的装置,其特征在于,The apparatus of claim 52 wherein,所述第一无线信道模型包括第一信道估计模型;The first wireless channel model includes a first channel estimation model;所述第一信道信息包括第一信道估计结果;The first channel information includes a first channel estimation result;所述第二信道信息包括参考信号。The second channel information includes a reference signal.
- 根据权利要求53或54所述的装置,其特征在于,Apparatus according to claim 53 or 54, characterized in that,所述第一无线信道模型包括第一信道状态信息CSI自编码模型,所述第二无线信道模型包括第二CSI自编码模型;The first wireless channel model includes a first channel state information CSI self-encoding model, and the second wireless channel model includes a second CSI self-encoding model;所述第一信道信息包括第一CSI恢复结果。The first channel information includes a first CSI recovery result.
- 根据权利要求48至56任一所述的装置,其特征在于,所述第二更新模块,用于在基于第一概率值确定所述第一无线信道模型满足更新条件的情况下,更新所述第一无线信道模型,所述第一概率值是通过所述第一判别器基于所述第一信道信息得到的。The device according to any one of claims 48 to 56, wherein the second update module is configured to update the A first wireless channel model, the first probability value is obtained by the first discriminator based on the first channel information.
- 根据权利要求57所述的装置,其特征在于,所述第一信道信息包括多个信道信息,所述第一概率值包括与所述多个信道信息一一对应的多个概率值;The device according to claim 57, wherein the first channel information includes a plurality of channel information, and the first probability value includes a plurality of probability values corresponding to the plurality of channel information one-to-one;所述更新条件包括如下条件中的至少一种:The update conditions include at least one of the following conditions:所述第一概率值中概率值低于第一阈值的占比高于预设值;The proportion of probability values lower than the first threshold among the first probability values is higher than a preset value;所述第一概率值中连续的x个概率值低于第二阈值,x为正整数;x consecutive probability values in the first probability value are lower than the second threshold, and x is a positive integer;根据所述第一概率值得到的概率值分布情况满足第一条件;The probability value distribution obtained according to the first probability value satisfies the first condition;基于所述第一概率值计算得到的评价数值达到第三阈值。The evaluation value calculated based on the first probability value reaches a third threshold.
- 根据权利要求57所述的装置,其特征在于,所述第一判别器包括至少两个子判别器,所述第一概率值包括至少两个子概率值。The apparatus according to claim 57, wherein the first discriminator comprises at least two sub-discriminators, and the first probability value comprises at least two sub-probability values.
- 根据权利要求59所述的装置,其特征在于,所述第二更新模块,用于在基于所述至少两个子概率值中的至少一个子概率值确定所述第一无线信道模型满足更新条件的情况下,更新所述第一无线信道模 型。The device according to claim 59, wherein the second update module is configured to determine that the first wireless channel model satisfies the update condition based on at least one sub-probability value of the at least two sub-probability values case, updating the first wireless channel model.
- 一种第一通信设备,其特征在于,所述第一通信设备包括:处理器;其中,A first communication device, characterized in that the first communication device includes: a processor; wherein,所述处理器,用于通过第一无线信道模型得到第一信道信息,所述第一无线信道模型为第一生成对抗网络中的生成器;The processor is configured to obtain first channel information through a first wireless channel model, where the first wireless channel model is a generator in a first generative confrontation network;所述处理器,用于在通过第一判别器基于所述第一信道信息确定所述第一无线信道模型满足更新条件的情况下,触发更新所述第一无线信道模型,所述第一判别器为所述第一生成对抗网络中的判别器。The processor is configured to trigger an update of the first radio channel model when the first discriminator determines that the first radio channel model meets an update condition based on the first channel information, and the first discriminator The discriminator is the discriminator in the first generative confrontation network.
- 一种第二通信设备,其特征在于,所述第二通信设备包括:处理器;其中,A second communication device, characterized in that the second communication device includes: a processor; wherein,所述处理器,用于在通过第一判别器基于第一信道信息确定第一无线信道模型满足更新条件的情况下,更新所述第一无线信道模型,所述第一信道信息是所述第一无线信道模型输出的,所述第一无线信道模型为第一生成对抗网络中的生成器,所述第一判别器为所述第一生成对抗网络中的判别器。The processor is configured to update the first wireless channel model when the first discriminator determines that the first wireless channel model satisfies an update condition based on the first channel information, the first channel information being the first channel information A wireless channel model is output, the first wireless channel model is a generator in the first generative adversarial network, and the first discriminator is a discriminator in the first generative adversarial network.
- 一种第一通信设备,其特征在于,所述第一通信设备包括:A first communication device, characterized in that the first communication device includes:处理器;processor;与所述处理器相连的收发器;a transceiver connected to the processor;用于存储所述处理器的可执行指令的存储器;memory for storing executable instructions of the processor;其中,所述处理器被配置为加载并执行所述可执行指令以实现如权利要求1至18中任一所述的无线信道模型的更新方法。Wherein, the processor is configured to load and execute the executable instructions to implement the wireless channel model updating method according to any one of claims 1 to 18.
- 一种第二通信设备,其特征在于,所述第二通信设备包括:A second communication device, characterized in that the second communication device includes:处理器;processor;与所述处理器相连的收发器;a transceiver connected to the processor;用于存储所述处理器的可执行指令的存储器;memory for storing executable instructions of the processor;其中,所述处理器被配置为加载并执行所述可执行指令以实现如权利要求19至30中任一所述的无线信道模型的更新方法。Wherein, the processor is configured to load and execute the executable instructions to implement the method for updating the wireless channel model according to any one of claims 19 to 30.
- 一种计算机可读存储介质,其特征在于,所述可读存储介质中存储有可执行指令,所述可执行指令由所述处理器加载并执行以实现如权利要求1至30中任一所述的无线信道模型的更新方法。A computer-readable storage medium, characterized in that executable instructions are stored in the readable storage medium, and the executable instructions are loaded and executed by the processor so as to implement any one of claims 1 to 30. The update method of the wireless channel model described above.
- 一种计算机程序产品,其特征在于,所述计算机程序产品中存储有可执行指令,所述可执行指令由所述处理器加载并执行以实现如权利要求1至30中任一所述的无线信道模型的更新方法。A computer program product, characterized in that executable instructions are stored in the computer program product, and the executable instructions are loaded and executed by the processor to implement the wireless communication system according to any one of claims 1 to 30. Update method for the channel model.
- 一种芯片,其特征在于,所述芯片用于实现如权利要求1至30中任一所述的无线信道模型的更新方法。A chip, characterized in that the chip is used to implement the method for updating a wireless channel model as claimed in any one of claims 1 to 30.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/CN2022/075697 WO2023150943A1 (en) | 2022-02-09 | 2022-02-09 | Method for updating wireless channel model, and apparatus, device and storage medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/CN2022/075697 WO2023150943A1 (en) | 2022-02-09 | 2022-02-09 | Method for updating wireless channel model, and apparatus, device and storage medium |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2023150943A1 true WO2023150943A1 (en) | 2023-08-17 |
Family
ID=87563392
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2022/075697 WO2023150943A1 (en) | 2022-02-09 | 2022-02-09 | Method for updating wireless channel model, and apparatus, device and storage medium |
Country Status (1)
Country | Link |
---|---|
WO (1) | WO2023150943A1 (en) |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111160128A (en) * | 2019-12-11 | 2020-05-15 | 中国资源卫星应用中心 | Remote sensing image processing method and system based on antagonistic neural network model |
CN111355675A (en) * | 2020-03-11 | 2020-06-30 | 南京航空航天大学 | Channel estimation enhancement method, device and system based on generation countermeasure network |
US11087215B1 (en) * | 2020-07-25 | 2021-08-10 | Sas Institute Inc. | Machine learning classification system |
CN113364528A (en) * | 2021-06-07 | 2021-09-07 | 中山大学 | Underwater wireless optical communication receiver for generating countermeasure network based on depth condition |
CN113381952A (en) * | 2021-06-09 | 2021-09-10 | 东南大学 | Multi-antenna system channel estimation method based on deep learning |
CN113381828A (en) * | 2021-06-08 | 2021-09-10 | 哈尔滨工业大学 | Sparse code multiple access random channel modeling method based on condition generation countermeasure network |
US20210295167A1 (en) * | 2020-03-23 | 2021-09-23 | Ansys, Inc. | Generative networks for physics based simulations |
-
2022
- 2022-02-09 WO PCT/CN2022/075697 patent/WO2023150943A1/en unknown
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111160128A (en) * | 2019-12-11 | 2020-05-15 | 中国资源卫星应用中心 | Remote sensing image processing method and system based on antagonistic neural network model |
CN111355675A (en) * | 2020-03-11 | 2020-06-30 | 南京航空航天大学 | Channel estimation enhancement method, device and system based on generation countermeasure network |
US20210295167A1 (en) * | 2020-03-23 | 2021-09-23 | Ansys, Inc. | Generative networks for physics based simulations |
US11087215B1 (en) * | 2020-07-25 | 2021-08-10 | Sas Institute Inc. | Machine learning classification system |
CN113364528A (en) * | 2021-06-07 | 2021-09-07 | 中山大学 | Underwater wireless optical communication receiver for generating countermeasure network based on depth condition |
CN113381828A (en) * | 2021-06-08 | 2021-09-10 | 哈尔滨工业大学 | Sparse code multiple access random channel modeling method based on condition generation countermeasure network |
CN113381952A (en) * | 2021-06-09 | 2021-09-10 | 东南大学 | Multi-antenna system channel estimation method based on deep learning |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Qin et al. | Deep learning in physical layer communications | |
Chaccour et al. | Less data, more knowledge: Building next generation semantic communication networks | |
EP4053754B1 (en) | Adaptive channel coding using machine-learned models | |
Mulvey et al. | Cell fault management using machine learning techniques | |
CN111062464B (en) | Power communication network reliability prediction and guarantee method and system based on deep learning | |
CN111242171B (en) | Model training and diagnosis prediction method and device for network faults and electronic equipment | |
Shi et al. | Machine learning for large-scale optimization in 6g wireless networks | |
CN101808244A (en) | Video transmission control method and system | |
CN111371534B (en) | Data retransmission method and device, electronic equipment and storage medium | |
CN102413378A (en) | Adaptive neural network-based lost packet recovery method in video transmission | |
CN114039870B (en) | Deep learning-based real-time bandwidth prediction method for video stream application in cellular network | |
Lan et al. | Progressive feature transmission for split classification at the wireless edge | |
WO2023150943A1 (en) | Method for updating wireless channel model, and apparatus, device and storage medium | |
CN116842440A (en) | Self-adaptive link switching method, system, equipment and medium based on context awareness | |
Zhang et al. | Blind recognition of channel coding based on CNN-BLSTM | |
US20220391731A1 (en) | Agent decision-making method and apparatus | |
CN113923099B (en) | Root cause positioning method for communication network fault and related equipment | |
Thomas et al. | Reasoning with the Theory of Mind for Pragmatic Semantic Communication | |
Yu et al. | Two-way semantic communications without feedback | |
Li et al. | Knowledge-Assisted Few-Shot Fault Diagnosis in Cellular Networks | |
Li et al. | Learning to Transmit with Provable Guarantees in Wireless Federated Learning | |
Yang et al. | Agent-driven Generative Semantic Communication for Remote Surveillance | |
CN118037048B (en) | Urban storm waterlogging disaster chain risk assessment system based on Bayesian network | |
WO2022257121A1 (en) | Communication method and device, and storage medium | |
CN117978247B (en) | Data compression method and device based on high-flux satellite Internet of things |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 22925305 Country of ref document: EP Kind code of ref document: A1 |