CN116882284A - Deep learning model construction method suitable for low-orbit satellite service - Google Patents
Deep learning model construction method suitable for low-orbit satellite service Download PDFInfo
- Publication number
- CN116882284A CN116882284A CN202310851909.7A CN202310851909A CN116882284A CN 116882284 A CN116882284 A CN 116882284A CN 202310851909 A CN202310851909 A CN 202310851909A CN 116882284 A CN116882284 A CN 116882284A
- Authority
- CN
- China
- Prior art keywords
- training
- deep learning
- service
- real
- low
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000013136 deep learning model Methods 0.000 title claims abstract description 25
- 238000010276 construction Methods 0.000 title claims abstract description 4
- 238000012549 training Methods 0.000 claims abstract description 51
- 238000000034 method Methods 0.000 claims abstract description 35
- 238000004364 calculation method Methods 0.000 claims abstract description 7
- 230000006870 function Effects 0.000 claims description 22
- 238000013528 artificial neural network Methods 0.000 claims description 17
- 238000004088 simulation Methods 0.000 claims description 15
- 210000002569 neuron Anatomy 0.000 claims description 14
- 238000009826 distribution Methods 0.000 claims description 12
- 230000008569 process Effects 0.000 claims description 11
- 230000004913 activation Effects 0.000 claims description 7
- 230000000694 effects Effects 0.000 claims description 4
- 238000011478 gradient descent method Methods 0.000 claims description 4
- 230000001537 neural effect Effects 0.000 claims description 2
- 238000010606 normalization Methods 0.000 claims description 2
- 230000003213 activating effect Effects 0.000 claims 1
- 238000013135 deep learning Methods 0.000 abstract description 5
- 238000004422 calculation algorithm Methods 0.000 description 7
- 238000004891 communication Methods 0.000 description 5
- 238000005516 engineering process Methods 0.000 description 5
- 238000005457 optimization Methods 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- 230000016273 neuron death Effects 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000013468 resource allocation Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000003860 storage Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 239000002699 waste material Substances 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- 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
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B7/00—Radio transmission systems, i.e. using radiation field
- H04B7/14—Relay systems
- H04B7/15—Active relay systems
- H04B7/185—Space-based or airborne stations; Stations for satellite systems
- H04B7/1851—Systems using a satellite or space-based relay
- H04B7/18519—Operations control, administration or maintenance
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/14—Network analysis or design
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/16—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks using machine learning or artificial intelligence
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/12—Timing analysis or timing optimisation
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- Signal Processing (AREA)
- Computer Networks & Wireless Communication (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Molecular Biology (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computing Systems (AREA)
- Medical Informatics (AREA)
- Databases & Information Systems (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- Astronomy & Astrophysics (AREA)
- Aviation & Aerospace Engineering (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention relates to the technical field of satellite mobile edge calculation, in particular to a deep learning model construction method suitable for low-orbit satellite service. The method solves the problems of low performance of the initial model, lack of pre-training data and high time cost for acquiring the pre-training data in a satellite network environment. The method realizes the centralized training of the model in an offline environment, and improves the online decision-making efficiency based on the model and the pre-training model, thereby improving the availability and the robustness of the satellite network deep learning application.
Description
Technical Field
The invention relates to the technical field of satellite mobile edge calculation, in particular to a deep learning model architecture method suitable for low-orbit satellite service.
Background
With the rise of 5G technology, the satellite network field has emerged with more and more services with large data volume (such as image remote sensing service). Traditional approaches to offloading services to ground gateway processing have become increasingly difficult to meet the delay requirements of the service due to the large amount of satellite-to-ground link bandwidth that is required. Satellite mobile edge computing techniques have therefore been proposed and are a hotspot for research. The technology applies edge computing technology to satellite communication systems to improve the performance and quality of service of the satellite communication systems. The method combines the advantages of wide area coverage, high-speed communication, wireless access and the like of a satellite communication system and the capabilities of data processing, storage, distribution and the like of edge calculation, and provides more efficient and reliable service for users.
In a satellite communication system, the dynamic scheduling and resource allocation of satellite mobile edge computing services can be realized by using a deep learning technology. Deep learning is a machine learning technology capable of processing a large amount of data, and can automatically learn and optimize a model, predict service demands for a period of time in the future by learning historical data and current states, or automatically learn and optimize a service scheduling policy according to current service demands and resource conditions. However, deep learning is highly dependent on large amounts of data that are independent and truly distributed for training. Untrained models often perform poorly to requirements, so a pre-trained model is necessary.
However, the information of the pre-training service is obtained from the execution of the service, and the large amount of training data means that a huge amount of running time is required, which is usually much longer than the time required for training, and thus, great waste is generated, which is not acceptable. Furthermore, the data of the pre-training model may have errors with the actually performed service information, which gradually increases with time, so how to eliminate this error with the real data is also a key problem.
Therefore, it is necessary to provide a deep learning model architecture method suitable for low-orbit satellite service to solve the above technical problems.
Disclosure of Invention
In order to solve the technical problems, the invention provides a deep learning model architecture method suitable for low-orbit satellite service.
The invention provides a deep learning model architecture method suitable for low-orbit satellite service, which comprises the following steps:
s1, setting equally configured virtual machines or real machine nodes based on hardware information of satellites, and constructing a simulation cluster;
s2, deploying various low-orbit satellite services with random numbers on a simulation cluster, collecting the execution time of each group of services, and repeating the process to obtain execution data as detailed as possible;
s3, arranging the data such as the resource configuration information, the current resource occupation information, the current execution service information, the service resource demand information to be allocated and the like of the nodes in the execution data into a vector form, and inputting the vector form into a deep neural network;
s4, training the deep neural network;
s5, introducing the deep neural network trained in the S4 into a simulation environment, and outputting estimated running time of the low-orbit satellite service as a time estimation function, so that offline training is carried out on the target deep learning model under the condition of no real service running;
s6, copying the deep learning model parameters which are subjected to multiple rounds and large sample offline training into a real environment, performing scheduling decision by the real environment according to the model, and storing execution data related to real service;
and S7, executing a data training time estimation function by using the real service every time a period of time passes, and improving the proportion of the real data in training, so that the simulation environment is always close to the real environment.
It should be noted that:
preferably, the training of the deep neural network in S4 comprises the following specific steps:
s41, designing a full connection layer with a proper number of layers according to the number of bits of an input vector;
s42, adding an activation function ReLU after each full connection layer, and referring to the activation function to determine whether the output of the neuron is activated or not, and adding nonlinear characteristics for the network to improve the expression capacity and fitting capacity of the network;
s43, adding a Dropout layer between every two full-connection layers, reducing the interdependence among neurons by randomly discarding part of neuron output, thereby reducing over-fitting and improving the robustness and robustness of the model;
s44, after each Dropout layer, adding a Layernormalization layer, and normalizing each sample in the characteristic dimension to avoid that the training weight of some prominent attributes in the input vector is overlarge so as to influence the overall training effect;
s45, using a mean square error as a loss function, and taking real execution time as a label training model;
s46, optimizing the model to reach a relative optimal value based on a random gradient descent method.
It should be noted that: let us assume that the input vector is x, a l Is a layer I neuron (a) 0 X), with output z of each layer l =W l *a l-1 +b l The method comprises the steps of carrying out a first treatment on the surface of the Wherein W is l A weight matrix for the layer I neurons;
preferably, the activation function ReLU in S42 is expressed as
It should be noted that: alpha is a small enough positive number, and the expression is set to be the above, so that the problem of 'neuron death' (that is, that a certain section of output is always for) can be solved;
preferably, the Dropout layer in S43 performs a neuronal rounding based on Bernoulli distribution, i.e. r-Bernoulli (p), y=r×h, where h is the output of the previous layer.
Preferably, in the step S44, the LayerNormalization layer normalizes the distribution of the input data to make the average value of the distribution of the input data be 0 and the standard deviation of the distribution of the input data be 1, so as to avoid disturbance of the input data with different probability distributions on the training process, where a specific formula is as follows:
preferably, the calculation formula of the mean square error in S45 is as follows:
where n is the number of samples, y i Is the true value of the i-th sample,is the predicted value of the i-th sample.
It should be noted that: the calculation method of the mean square error is that firstly, the squares of the difference between the predicted value and the true value of each sample are added up, then the average value is calculated, and the smaller the obtained value is, the better the prediction capability of the model is;
preferably, the formula of the S46 random gradient descent is:
wherein θ is t Representing model parameters epsilon after the t-th round of iteration t In order for the rate of learning to be high,gradient with respect to the ith sample for the loss function.
Compared with the related art, the deep learning model architecture method suitable for the low-orbit satellite service has the following beneficial effects:
1. according to the invention, the time attribute in the real application execution process is fitted by using the deep neural network, so that a simulated simulation environment close to the real environment is constructed, and test and pre-training data are generated, thereby being capable of updating the deep learning model in the satellite network application efficiently, synchronously and immediately. The method solves the problems of low performance of the initial model, lack of pre-training data and high time cost for acquiring the pre-training data in a satellite network environment. The method realizes the centralized training of the model in an offline environment, and improves the online decision-making efficiency based on the model and the pre-training model, thereby improving the availability and the robustness of the satellite network deep learning application;
drawings
FIG. 1 is a schematic diagram of a deep neural network training process according to the present invention;
FIG. 2 is a schematic flow chart of a deep learning model architecture method suitable for low-orbit satellite service;
FIG. 3 is a graph comparing service completion rates under offline and real environment deployment decisions based on image recognition applications;
FIG. 4 is a comparison of service scores for offline and real-world environment deployment decisions based on image recognition applications.
Detailed Description
In order to make the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the described embodiments are some embodiments of the present invention, but not all embodiments of the present invention. The components of the embodiments of the present invention generally described and illustrated in the figures herein may be arranged and designed in a wide variety of different configurations.
Thus, the following detailed description of the embodiments of the invention, as presented in the figures, is not intended to limit the scope of the invention, as claimed, but is merely representative of selected embodiments of the invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
It should be noted that: like reference numerals and letters denote like items in the following figures, and thus once an item is defined in one figure, no further definition or explanation thereof is necessary in the following figures.
Specific implementations of the invention are described in detail below in connection with specific embodiments.
Referring to fig. 1 to 2, the present invention provides a deep learning model suitable for low-orbit satellite service
A kind of electronic device with a display unit
A method of architecture, comprising:
s1, setting equally configured virtual machines or real machine nodes based on hardware information of satellites;
s2, deploying various low-orbit satellite services with random numbers to virtual machines or real machine nodes, collecting execution time of each service, and repeating the process to obtain execution data as detailed as possible;
s3, arranging the data such as the resource configuration information, the current resource occupation information, the current execution service information, the service resource demand information to be allocated and the like of the nodes in the execution data into a vector form, and inputting the vector form into a deep neural network;
s4, training the deep neural network;
s5, introducing the deep neural network trained in the S4 into a simulation environment, and outputting estimated running time of the low-orbit satellite service as a time estimation function, so that offline training is carried out on the target deep learning model under the condition of no real service running;
s6, copying the deep learning model parameters which are subjected to multiple rounds and large sample offline training into a real environment, performing scheduling decision by the real environment according to the model, and storing execution data related to real service;
and S7, executing a data training time estimation function by using the real service every time a period of time passes, and improving the proportion of the real data in training, so that the simulation environment is always close to the real environment.
It should be noted that: through the virtual machine and the real machine node, a simulation cluster is constructed, various low-orbit satellite services with random numbers can be deployed on the real machine node, a simulation environment is deployed on the virtual machine, and offline training is carried out on the deep learning model on the virtual machine.
Preferably, the training of the deep neural network in S4 comprises the following specific steps:
s41, designing a full connection layer with a proper number of layers according to the number of bits of an input vector;
s42, adding an activation function ReLU after each full connection layer, and referring to the activation function to determine whether the output of the neuron is activated or not, and adding nonlinear characteristics for the network to improve the expression capacity and fitting capacity of the network;
s43, adding a Dropout layer between every two full-connection layers, reducing the interdependence among neurons by randomly discarding part of neuron output, thereby reducing over-fitting and improving the robustness and robustness of the model;
s44, after each Dropout layer, adding a Layernormalization layer, and normalizing each sample in the characteristic dimension to avoid that the training weight of some prominent attributes in the input vector is overlarge so as to influence the overall training effect;
s45, using a mean square error as a loss function, and taking real execution time as a label training model;
s46, optimizing the model to reach a relative optimal value based on a random gradient descent method.
And S4, training the deep neural network, wherein the modeling method of the specific steps is as follows:
1. let us assume that the input vector is x, a l Is a layer I neuron (a) 0 X), with output z of each layer l =W l *a l-1 +b l The method comprises the steps of carrying out a first treatment on the surface of the Wherein W is l Is the weight matrix of the layer I neurons.
2. The expression of the ReLU function is
To solve the problem of "neuronal death" (i.e. a certain segment of output is always true), we introduced a leak ReLU, rewritten the above formula asWherein α is a sufficiently small positive number;
3. the Dropout layer trades off neurons based on Bernoulli distribution, i.e., there are r-Bernoulli (p), y=r h, where h is the output of the previous layer.
4. The Layer Normalization (LN) layer normalizes the distribution of input data to a mean value of 0 standard deviation of 1,
thereby avoiding disturbance of input data with different probability distributions to the training process, and the specific formula is as follows:
5. the mean square error calculating method is to sum up the squares of the differences between the predicted value and the true value of each sample, and then average the values, and the smaller the obtained value is, the better the prediction capability of the model is.
The formula is as follows:
where n is the number of samples, y i Is the true value of the i-th sample,pre-processing for the ith sample
And (5) measuring values.
6. After the difference between the input and the true value is obtained through the loss function, a random gradient descent method is utilized to counter-propagate and train a random gradient descent formula of the deep neural network, wherein the formula is as follows:
wherein θ is t Representing model parameters epsilon after the t-th round of iteration t In order for the rate of learning to be high,as a loss function
Gradient with respect to the ith sample.
Since the random gradient descent uses only one sample gradient to update model parameters, the change in update direction is more random, possibly resulting in a relatively unstable update process, and we have chosen a smaller initial learning rate (ε) t =1e-3) to ensure convergence. Furthermore, we incorporate the optimization algorithm Adam, since random sampling of samples by random gradient descent may lead to instability of the training process.
Adam (Adaptive Moment Estimation) is an adaptive learning rate optimization algorithm, and the basic idea is to combine the advantages of two optimization algorithms, namely Momentum and RMSProp, so that the stability of parameter updating can be ensured, and the learning rate can be adaptively adjusted. The update formula is as follows:
m t =β 1 m t-1 +(1-β 1 )g t
v t =β 2 m t-1 +(1-β 2 )(g t ) 2
wherein m is t And v t Respectively, are the index weighted average of the gradient and the square of the gradient before the t-th round of iteration, beta 1 And beta 2 The attenuation coefficient, which is the average value, is generally 0.9 and 0.999, respectively.And->For m t And v t Offset correction is performed to eliminate the effects of the previous rounds of iterations on the mean and variance.
The working principle of the deep learning model architecture method suitable for the low-orbit satellite service provided by the invention is as follows: the time to perform using deep neural network fitting can thus greatly reduce the time required for pre-training. Specifically, the execution time may be taken as an output, the characteristics of the service as an input, and then the execution time of the service may be fitted by training a deep neural network model. In this way, a pre-trained model can be obtained without relying on large-scale data sets, thereby greatly reducing the pre-training time.
FIG. 3 is a graph comparing scores of image recognition algorithms for different numbers of services in a real environment and a simulated environment, the former being the simulated environment and the latter being the real environment.
The similarity scores achieved in the simulated environment and the real environment can be pre-trained based on the algorithm of the present invention. The minimum difference between the two is not more than 1% when the number of services is small; in the case of a large number of services, the maximum difference between the two is 26%. This is because the real environment is more hardware limited than the simulation environment, and thus it is difficult to load excessive services, thereby lowering the score.
Fig. 4 is a comparison diagram of the completion of the image recognition algorithm under different service numbers in a real environment and a simulation environment, the former is the simulation environment, and the latter is the real environment.
The service completed in the tolerant time is considered to be completed, and the graph shows that the model pre-trained by the invention has very close performance in the real environment and the simulation environment, and the maximum difference is not more than 2.3%, so that the model pre-trained by the algorithm can be described to be suitable for the real environment.
The circuits and control involved in the present invention are all of the prior art, and are not described in detail herein. The foregoing description is only illustrative of the present invention and is not intended to limit the scope of the invention, and all equivalent structures or equivalent processes or direct or indirect application in other related technical fields are included in the scope of the present invention.
Claims (7)
1. A deep learning model construction method suitable for low orbit satellite service is characterized in that,
tool with
The method comprises the following steps:
s1, setting equally configured virtual machines or real machine nodes based on hardware information of satellites;
s2, deploying various low-orbit satellite services with random quantity to virtual machines or real machine nodes, and collecting each satellite
The execution time of the group service and repeating the process to acquire execution data as detailed as possible;
s3, arranging data such as resource configuration information, current resource occupation information, current execution service information, service resource demand information to be allocated and the like of nodes in the execution data into a vector form, and inputting the vector form into a deep neural network;
s4, training the deep neural network;
s5, introducing the deep neural network trained in the S4 into a simulation environment, and outputting estimated running time of the low-orbit satellite service as a time estimation function, so that offline training is carried out on the target deep learning model under the condition of no real service running;
s6, copying the deep learning model parameters which are subjected to multiple rounds and large sample offline training into a real environment, performing scheduling decision by the real environment according to the model, and storing execution data related to real service;
and S7, executing a data training time estimation function by using the real service every time a period of time passes, and improving the proportion of the real data in training, so that the simulation environment is always close to the real environment.
2. The deep learning model architecture method for low-orbit satellite service according to claim 1, wherein the training of the deep neural network in S4 comprises the following specific steps:
s41, designing a full connection layer with a proper number of layers according to the number of bits of an input vector;
s42, adding an activation function ReLU after each full connection layer, and referring to the activation function to determine whether the output of the neuron is activated or not, and adding nonlinear characteristics for the network to improve the expression capacity and fitting capacity of the network;
s43, adding a Dropout layer between every two full-connection layers, reducing the interdependence among neurons by randomly discarding part of neuron output, thereby reducing over-fitting and improving the robustness and robustness of the model;
s44, after each Dropout layer, adding a Layernormalization layer, and normalizing each sample in the characteristic dimension to avoid that the training weight of some prominent attributes in the input vector is overlarge so as to influence the overall training effect;
s45, using a mean square error as a loss function, and taking real execution time as a label training model;
s46, optimizing the model to reach a relative optimal value based on a random gradient descent method.
3. The method of claim 2, wherein the activating function ReLU in S42 is expressed as
4. The deep learning model architecture method for low-orbit satellite service according to claim 2, wherein the Dropout layer in S43 performs a neuronal rejection based on Bernoulli distribution, i.e. r-Bernoulli (p), y=r×h, where h is the output of the previous layer.
5. The deep learning model architecture method for low-orbit satellite service according to claim 2, wherein the step S44 is characterized in that the layer Layer Normalization normalizes the distribution of the input data to make the mean value of the input data be 0 and the standard deviation be 1, so as to avoid disturbance of the input data with different probability distributions to the training process, and the specific formula is as follows:
6. the deep learning model architecture method for low-orbit satellite service according to claim 2, wherein the calculation formula of the mean square error in S45 is as follows:
where n is the number of samples, y i Is the true value of the i-th sample,is the predicted value of the i-th sample.
7. The deep learning model architecture method for low-orbit satellite service according to claim 2, wherein the formula of S46 random gradient descent is:
wherein θ is t Representing model parameters epsilon after the t-th round of iteration t In order for the rate of learning to be high,gradient with respect to the ith sample for the loss function.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310851909.7A CN116882284A (en) | 2023-07-12 | 2023-07-12 | Deep learning model construction method suitable for low-orbit satellite service |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310851909.7A CN116882284A (en) | 2023-07-12 | 2023-07-12 | Deep learning model construction method suitable for low-orbit satellite service |
Publications (1)
Publication Number | Publication Date |
---|---|
CN116882284A true CN116882284A (en) | 2023-10-13 |
Family
ID=88261618
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202310851909.7A Pending CN116882284A (en) | 2023-07-12 | 2023-07-12 | Deep learning model construction method suitable for low-orbit satellite service |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN116882284A (en) |
-
2023
- 2023-07-12 CN CN202310851909.7A patent/CN116882284A/en active Pending
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111556461B (en) | Vehicle-mounted edge network task distribution and unloading method based on deep Q network | |
EP3805999A1 (en) | Resource-aware automatic machine learning system | |
US20180046919A1 (en) | Multi-iteration compression for deep neural networks | |
CN112351503A (en) | Task prediction-based multi-unmanned-aerial-vehicle-assisted edge computing resource allocation method | |
US11042802B2 (en) | System and method for hierarchically building predictive analytic models on a dataset | |
WO2021036414A1 (en) | Co-channel interference prediction method for satellite-to-ground downlink under low earth orbit satellite constellation | |
CN108805193B (en) | Electric power missing data filling method based on hybrid strategy | |
CN107103359A (en) | The online Reliability Prediction Method of big service system based on convolutional neural networks | |
CN115686846B (en) | Container cluster online deployment method integrating graph neural network and reinforcement learning in edge calculation | |
CN112765894B (en) | K-LSTM-based aluminum electrolysis cell state prediction method | |
CN112766603A (en) | Traffic flow prediction method, system, computer device and storage medium | |
CN112183742A (en) | Neural network hybrid quantization method based on progressive quantization and Hessian information | |
Gil et al. | Quantization-aware pruning criterion for industrial applications | |
CN114118567A (en) | Power service bandwidth prediction method based on dual-channel fusion network | |
US11914672B2 (en) | Method of neural architecture search using continuous action reinforcement learning | |
CN115564155A (en) | Distributed wind turbine generator power prediction method and related equipment | |
CN114936708A (en) | Fault diagnosis optimization method based on edge cloud collaborative task unloading and electronic equipment | |
CN113408610B (en) | Image identification method based on adaptive matrix iteration extreme learning machine | |
Urgun et al. | Composite power system reliability evaluation using importance sampling and convolutional neural networks | |
CN117523291A (en) | Image classification method based on federal knowledge distillation and ensemble learning | |
CN116882284A (en) | Deep learning model construction method suitable for low-orbit satellite service | |
KR20080078292A (en) | Domain density description based incremental pattern classification method | |
CN111079902A (en) | Decomposition fuzzy system optimization method and device based on neural network | |
CN110768825A (en) | Service flow prediction method based on network big data analysis | |
CN116303386A (en) | Intelligent interpolation method and system for missing data based on relational graph |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |