CN116319210A

CN116319210A - Signal lightweight automatic modulation recognition method and system based on deep learning

Info

Publication number: CN116319210A
Application number: CN202310291803.6A
Authority: CN
Inventors: 易冬; 胡涛; 王书; 刘凯越; 李汀立; 张靖志; 牛朝阳; 吴迪; 成凯鑫
Original assignee: Information Engineering University of PLA Strategic Support Force
Current assignee: Information Engineering University of PLA Strategic Support Force
Priority date: 2023-03-23
Filing date: 2023-03-23
Publication date: 2023-06-23

Abstract

The invention relates to the technical field of signal processing, in particular to a signal lightweight automatic modulation recognition method and system based on deep learning, which acquire a sample data set containing a plurality of digital modulation modes and a plurality of analog modulation modes by simulating a real communication environment modulation information sequence; constructing a lightweight dense convolution long-short time memory network structure based on a lightweight convolution network, and training and optimizing the lightweight dense convolution long-short time memory network structure by utilizing a sample data set; the signal modulation recognition model is built based on the light-weight dense convolution long-short time memory network structure after training and optimization; inputting the target signal to be identified into a signal modulation identification model, and acquiring a modulation mode of the target signal to be identified by using the signal modulation identification model. The invention solves the problems of high complexity, numerous parameters, huge model and the like in the existing automatic modulation identification, and meets the deployment requirement on resource-limited equipment through the design of a lightweight model.

Description

Signal lightweight automatic modulation recognition method and system based on deep learning

Technical Field

The invention relates to the technical field of signal processing, in particular to a signal lightweight automatic modulation recognition method and system based on deep learning.

Background

Automatic Modulation Recognition (AMR) can acquire a modulation pattern of a signal, which is a precondition for completing signal demodulation and thus acquiring information. Automatic modulation identification is a key technology in modern communication systems, and is widely applied, such as spectrum interference detection, spectrum sensing, cognitive radio and the like. Researchers have conducted a great deal of research in the field of automatic modulation recognition, and have proposed various automatic modulation recognition methods. The current automatic modulation recognition method of signals can be roughly classified into an automatic modulation recognition method (LB-AMR) based on likelihood ratio, an automatic modulation recognition method (FB-AMR) based on feature extraction, and an automatic modulation recognition method (DL-AMR) based on deep learning. LB-AMR is optimal in the sense of Bayesian estimation, but it relies heavily on a priori knowledge and parameter estimation, and the algorithm complexity is high. The FB-AMR extracts various manual characteristics such as instantaneous signal amplitude, phase and frequency, constellation diagram, time-frequency distribution characteristics, high-order accumulation quantity, cyclic spectrum and the like through an expert system, and then applies algorithms such as an artificial neural structure, a support vector machine, a decision tree and the like to a classification process so as to improve AMR performance. However, the extraction of the manual feature of the differentiation of the FB-AMR requires extensive domain knowledge, and the quality of the feature extraction directly determines the performance of the FB-AMR. DL-AMR is better than FB-AMR and LB-AMR to some extent.

However, the existing deep learning algorithm is not designed for AMR, and some algorithm frames directly referred to from the fields of image recognition, voice recognition, and the like achieve a certain effect in the AMR field, but the algorithm parameters are often numerous, and the model is huge and complex. On platforms with less available resources or precious resources, such as satellite platforms, the computing power of the on-board computer is very different from the storage space relative to the ground computer due to constraints of volume, mass, power consumption, and environmental factors such as space radiation, extreme temperatures, and maintenance difficulties. Therefore, a light-weight and low-complexity DL-AMR method has been studied, and how to enable DL-AMR to be deployed in a resource-constrained device as a development direction in a signal modulation recognition direction by reducing a model size or accelerating a calculation time while ensuring recognition accuracy.

Disclosure of Invention

Therefore, the invention provides a signal lightweight automatic modulation recognition method and system based on deep learning, which solve the problems of high complexity, numerous parameters, huge model and the like in the existing automatic modulation recognition, and meet the deployment requirement on resource-limited equipment through lightweight model design.

According to the design scheme provided by the invention, the signal lightweight automatic modulation identification method based on deep learning comprises the following steps:

acquiring a sample data set containing a plurality of digital modulation modes and a plurality of analog modulation modes by simulating a real communication environment modulation information sequence;

constructing a lightweight dense convolution long-short time memory network structure based on a lightweight convolution network, and training and optimizing the lightweight dense convolution long-short time memory network structure by utilizing a sample data set; the signal modulation recognition model is built based on the light-weight dense convolution long-short time memory network structure after training and optimization;

inputting the target signal to be identified into a signal modulation identification model, and acquiring a modulation mode of the target signal to be identified by using the signal modulation identification model.

As the signal lightweight automatic modulation recognition method based on deep learning, the invention further obtains a sample data set containing a plurality of digital modulation modes and a plurality of analog modulation modes by simulating a real communication environment modulation information sequence, and comprises the following steps:

firstly, randomly generating a 01 bit information sequence of a radio by utilizing a random function; the method comprises the steps of carrying out a first treatment on the surface of the

Then, the real communication environment modulation information sequence is simulated using the 01-bit information sequence, and N digital modulation signals and M analog modulation signals are generated to construct a sample data set using the N digital modulation signals and the M analog modulation signals, wherein M, N is an integer greater than 1, respectively.

As the signal lightweight automatic modulation recognition method based on deep learning, the invention further utilizes a 01-bit information sequence to simulate a real communication environment modulation information sequence, and adds real communication environment influence parameters in the modulation and sampling processes of the information sequence to obtain an IQ sampling sequence of a simulated modulation signal, wherein the real communication environment influence parameters comprise but are not limited to: additive gaussian noise, multipath fading, sample rate offset, and center frequency offset.

As the signal lightweight automatic modulation recognition method based on deep learning, the invention further utilizes N digital modulation signals and M analog modulation signals to form a sample data set, and for the information points of each modulation signal, K information points are continuously collected at each time by taking the information points with a preset threshold number as sampling intervals to form a signal sample; each modulation signal samples a preset number of samples, and the samples sampled by all the modulation signals are combined to construct a signal sample data set, wherein the preset number is set in units of ten thousands.

As the signal lightweight automatic modulation recognition method based on deep learning, the invention further discloses a lightweight dense convolution long-short time memory network structure constructed based on a lightweight convolution network, which comprises the following steps: the method comprises the steps of inputting data subjected to standardization processing to a batch normalization processing unit, carrying out convolution and splicing operation on the standardized data, and extracting time sequence features of the spliced data by using a long-short-time memory network and classifying the time sequence features by activating a function.

As the signal lightweight automatic modulation recognition method based on deep learning, the feature extraction and classification unit is formed by connecting two long-short-time memory network layers and a full-connection layer, wherein the two long-short-time memory network layers adopt different numbers of hidden units, and the full-connection layer activation function adopts a Softmax function.

When the method is used for carrying out training optimization on the lightweight dense convolution long-short-time memory network structure by utilizing the sample data set, an absolute cross entropy loss function is set as a target loss function in the training process, an Adam optimizer is selected for carrying out optimization on the network, and the maximum iteration round or early shutdown is set as an iteration termination condition for training optimization.

Further, based on the method, the invention also provides a signal lightweight automatic modulation recognition system based on deep learning, which comprises the following steps: a sample construction module, a model training module and a target recognition module, wherein,

the sample construction module is used for acquiring a sample data set containing a plurality of digital modulation modes and a plurality of analog modulation modes by simulating a real communication environment modulation information sequence;

the model training module is used for constructing a lightweight dense convolution long-short time memory network structure based on the lightweight convolution network, and training and optimizing the lightweight dense convolution long-short time memory network structure by utilizing the sample data set; the signal modulation recognition model is built based on the light-weight dense convolution long-short time memory network structure after training and optimization;

the target recognition module is used for inputting the target signal to be recognized into the signal modulation recognition model, and the modulation mode of the target signal to be recognized is obtained by utilizing the signal modulation recognition model.

The invention has the beneficial effects that:

the invention generates 8 existing digital modulation signals and 2 analog modulation signals by simulating a real communication environment modulation information sequence so as to construct a sample data set for model training optimization; the method is characterized in that a light-weight dense convolution long-short-time memory network model for signal modulation recognition is constructed based on a light-weight convolution neural network structure, training and optimization are carried out by utilizing a sample data set, a target signal is recognized by utilizing the model after training and optimization, the problems of high complexity, numerous parameters, huge model and the like in the existing automatic modulation recognition are solved, the recognition rate is ensured through the light-weight design of the model, the deployment on resource-limited equipment such as a satellite platform can be met, and the method is convenient for practical scene application.

Description of the drawings:

FIG. 1 is a schematic illustration of a signal lightweight automatic modulation recognition flow based on deep learning in an embodiment;

FIG. 2 is a schematic diagram of a signal modulation recognition model construction flow in an embodiment;

FIG. 3 is a diagram showing a change in validation loss during a training process of a network model in an embodiment;

fig. 4 is a schematic diagram of a signal modulation recognition rate result in an embodiment.

The specific embodiment is as follows:

the present invention will be described in further detail with reference to the drawings and the technical scheme, in order to make the objects, technical schemes and advantages of the present invention more apparent.

Referring to fig. 1, the embodiment of the invention provides a signal lightweight automatic modulation identification method based on deep learning, which comprises the following steps:

s101, acquiring a sample data set containing a plurality of digital modulation modes and a plurality of analog modulation modes through simulating a real communication environment modulation information sequence.

Specifically, a 01 bit information sequence of a radio may be first randomly generated using a random function; then, the real communication environment modulation information sequence is simulated using the 01-bit information sequence, and N digital modulation signals and M analog modulation signals are generated to construct a sample data set using the N digital modulation signals and the M analog modulation signals, wherein M, N is an integer greater than 1, respectively.

In the method, a 01 bit information sequence is utilized to simulate a real communication environment modulation information sequence, and real communication environment influence parameters are added in the process of modulating and sampling the information sequence to obtain an IQ sampling sequence of a simulated modulation signal, wherein the real communication environment influence parameters include but are not limited to: additive gaussian noise, multipath fading, sample rate offset, and center frequency offset.

And a 01 bit sequence is randomly generated by utilizing a random function, so that the randomness of the information content is ensured, and the influence of the signal content on the signal modulation identification can be removed. The information sequence is modulated in the simulated real communication environment, the quadrature in-phase IQ sequence is obtained by sampling, and the influence of additive Gaussian noise, multipath fading, sampling rate offset and center frequency offset is added in the modulation and sampling process of the information sequence, so that the method is similar to the actual environment. The partial parameters are as follows: the modulation rate is 25KBaud/s, the sampling rate is 200KHz, the standard deviation of each sample point in the sampling rate drifting process is 1Hz, the maximum sampling rate is shifted by 50Hz, the maximum Doppler frequency used in fading simulation is 1Hz, the time delay vector [1,0.8,0.3], the signal to noise ratio is between-20 dB and 18dB at intervals of 2dB, and the number of sine waves used in frequency selective fading simulation is 8. The IQ sample sequence of the modulated signal, i.e. two values per information point, 8 information points per symbol, can be obtained most.

It should be noted that the specific values of the parameter settings are not within the protection range of the scheme, and the specific values of the parameters can be adjusted according to the actual experimental environment. The M/N can respectively represent 8 digital modulation signals and 2 analog modulation signals, namely, the 8 digital modulation modes of 8PSK,BPSK,CPFSK,GFSK,PAM4, 16QAM,64QAM and the two analog modulation modes of AM-DSB and WBFM.

As a preferred embodiment, further, a sample data set is constructed by using N digital modulation signals and M analog modulation signals, and for information points of each modulation signal, K information points are continuously collected at each time with a preset threshold number of information points as sampling intervals to form a signal sample; each modulation signal samples a preset number of samples, and the samples sampled by all the modulation signals are combined to construct a signal sample data set, wherein the preset number is set in units of ten thousands.

The information points of each modulation signal take 8 information points as sampling intervals, and 128 information points are continuously collected each time to form a signal sample. Each modulated signal acquired 12 ten thousand samples, all of which were grouped into a signal sample set. 60% of each type of modulation signal is extracted from the generated sample set to form a training sample set, 20% of the rest 40% of modulation signals are extracted to form a verification sample set, and the final 20% of the whole sample set is taken as a test sample set.

Each modulated signal may take 12 ten thousand samples, all of which constitute a signal sample set. For each modulation signal, signals are collected at 2dB intervals before-20 dB-18 dB, 6000 samples are collected under each signal-to-noise ratio, and 12 ten thousand samples are collected for each modulation signal. All the acquired signals are formed into a signal sample set, and 120 ten thousand samples are taken in total.

From the generated sample set, each type of modulation signal can be randomly extracted to form a training sample set, 20% of the rest 40% of modulation signals are randomly extracted to form a verification sample set, and the last 20% of the whole sample set is taken as a test sample set. For 6000 signal samples under each signal-to-noise ratio under each modulation pattern, firstly, randomly extracting 60% to be added into a training set, then randomly extracting 20% from the remaining 40% to be added into a verification sample set, and finally, adding the remaining 20% into a test sample set. Referring to fig. 2, the training sample set, the verification sample set and the test sample set are used to respectively perform training verification and test on the network model, so as to obtain a model structure finally used for modulating and identifying the target signal through training tuning.

It should be noted that the above specific values may be set according to experience values and/or actual use environments, and other specific values may be selected by those skilled in the art.

S102, constructing a lightweight dense convolution long-short time memory network structure based on a lightweight convolution network, and training and optimizing the lightweight dense convolution long-short time memory network structure by utilizing a sample data set; and a signal modulation recognition model is established based on the light-weight dense convolution long-short time memory network structure after training and optimization.

Specifically, a lightweight dense convolution long-short time memory network structure constructed based on a lightweight convolution network comprises: the method comprises the steps of inputting data subjected to standardization processing to a batch normalization processing unit, carrying out convolution and splicing operation on the standardized data, and extracting time sequence features of the spliced data by using a long-short-time memory network and classifying the time sequence features by activating a function.

In the data input batch normalization unit, an IQ sequence is input into a network in a one-dimensional time sequence format, and a Batch Normalization (BN) layer is arranged at the head of the network so as to normalize input signals, thereby solving the model training difficulty, preventing gradient elimination and internal data distribution offset phenomena and enhancing the robustness of the network to various received signals; the BN layer first calculates the mean and variance of the training data and then transforms the data to fit a standard normal distribution with a mean of 0 and a standard deviation of 1.

The dense connection convolution unit may be designed to contain 5 convolution units (convolution unit 1, convolution unit 2, convolution unit 3, convolution unit 4, convolution unit 5) and 5 splice layers (Concate 1, concate 2, concate 3, concate 4, concate 5). The convolution unit is composed of a convolution layer, a BN layer and a ReLU activation function in sequence, and nonlinear mapping capability is improved. The stitching layer is used for stitching the input according to the dimension.

The connection structure of the densely connected convolution units can be designed as follows: BN layer→convolution unit 1→conccate 1, BN layer→conccate 1; concate 1→convolution unit 2→Concate 2, concate 1→Concate 2;

concate 2→convolution unit 3→Concate 2→Concate 3; concate 3→convolution unit 4→Concate 4, concate 3→Concate 4; concate 4. Fwdarw. Convolution element 5 →

Concate 5; BN layer → Concatenate5

The feature extraction and classification unit can be formed by sequentially connecting two long and short time memory layers (LSTM 1 and LSTM 2) and a full-connection output layer. Wherein the activation function of the fully connected output layer is Softmax.

Parameters of 6 layers of convolution layers and parameters of 2 layers of long-short time memory layers in a lightweight intensive convolution long-short time memory network, a target loss function, an optimizer, an initial learning rate, a batch training size, training rounds and an early stopping mechanism can be set firstly; and training and optimizing the network model after parameter initialization by using the sample data set.

Specific parameters of the 6-layer convolution layer and the 2-layer long-short-time memory layer can be set as follows:

the number of convolution kernels of the one-dimensional convolution layer in the convolution unit 1 is 4, and the convolution kernel size is 3.

The number of convolution kernels of the one-dimensional convolution layer in the convolution unit 2 is 4, and the convolution kernel size is 3.

The number of convolution kernels of the one-dimensional convolution layer in the convolution unit 3 is 4, and the convolution kernel size is 5.

The number of convolution kernels of the one-dimensional convolution layer in the convolution unit 4 is 4 and the convolution kernel size is 7.

The number of convolution kernels of the one-dimensional convolution layer in the convolution unit 5 is 8, and the convolution kernel size is 1.

The number of hidden units of LSTM1 is 32.

The number of hidden units of LSTM2 is 10.

The number of convolution kernels of the full connection layer is 11, corresponding to the number of output categories.

When training a light-weight dense convolution long-short time memory network by using a training set, selecting an Adam optimizer to optimize the network, setting the initial learning rate to be 0.001, training 512 samples in each batch, and setting the maximum training round of the whole training samples to be 200. The model is validated using the validation set for each training round, with the validation loss as a reference, and the training of the model is stopped when the validation loss does not decrease after 50 durations.

Inputting the test sample set into a trained light-weight dense convolution long-short time memory network to obtain a recognition result, comparing the recognition result with a real category, and counting the recognition accuracy; and simultaneously recording the reasoning time of the light-weight dense convolution long-short-time memory network to obtain the sample reasoning speed so as to perform optimization processing on the network model parameters.

In model training optimization, the arrangement sequence of all samples in training samples can be disordered, the training samples and verification samples are input into a light-weight dense convolution long-short-time memory network model, a light-weight dense convolution long-short-time memory network is trained, and when the maximum turn of network training is achieved or the condition of early shutdown is met, the training process of the neural network is completed, and a trained light-weight dense convolution long-short-time memory network model is obtained.

S103, inputting the target signal to be identified into a signal modulation identification model, and acquiring a modulation mode of the target signal to be identified by using the signal modulation identification model.

The high-efficiency and high-precision recognition of various modulation signals is completed by using the signal modulation recognition model after training and optimization, the overall performance of signal modulation recognition is improved, and scheme deployment is facilitated on application equipment.

To verify the validity of this protocol, the following is further explained in connection with experimental data:

the simulation experiment is realized on NVIDIA Quadro RTX 6000 and Keras2.6.0Tensorflow-GPU2.4.0 platforms, and the simulation experiment of the invention and the generation of the modulation signal and the light-weight dense convolution long-short-time memory network is completed. And (3) completing an experiment by using the steps shown in fig. 2, and obtaining the verification loss change trend in the training process of the light-weight dense convolution long-short time memory network so as to verify the signal recognition rate and the model reasoning speed of the scheme.

FIG. 3 illustrates the change in validation loss during training of a lightweight dense convolution long and short time memory network; as can be seen from fig. 3, the verification loss decreases, converges and stabilizes, which indicates that the training effect of the simulation experiment becomes better as the training frequency increases. Fig. 4 is a graph of simulation experiment recognition rate results of the scheme, and the graph shows that the recognition rate is gradually increased and stabilized along with the increase of the signal-to-noise ratio, and the highest recognition rate can reach 93.7%; at the same time, the model inference speed was recorded as 0.024 milliseconds per sample inference time.

According to the simulation experiment, the scheme can efficiently and accurately complete the automatic modulation recognition task aiming at the automatic recognition of the modulation signal, and the feasibility of the scheme is further verified.

The relative steps, numerical expressions and numerical values of the components and steps set forth in these embodiments do not limit the scope of the present invention unless it is specifically stated otherwise.

In the present specification, each embodiment is described in a progressive manner, and each embodiment is mainly described in a different point from other embodiments, and identical and similar parts between the embodiments are all enough to refer to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant points refer to the description of the method section.

The elements and method steps of the examples described in connection with the embodiments disclosed herein may be embodied in electronic hardware, computer software, or a combination thereof, and the elements and steps of the examples have been generally described in terms of functionality in the foregoing description to clearly illustrate the interchangeability of hardware and software. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the solution. Those of ordinary skill in the art may implement the described functionality using different methods for each particular application, but such implementation is not considered to be beyond the scope of the present invention.

Those of ordinary skill in the art will appreciate that all or a portion of the steps in the above methods may be performed by a program that instructs associated hardware, and that the program may be stored on a computer readable storage medium, such as: read-only memory, magnetic or optical disk, etc. Alternatively, all or part of the steps of the above embodiments may be implemented using one or more integrated circuits, and accordingly, each module/unit in the above embodiments may be implemented in hardware or may be implemented in a software functional module. The present invention is not limited to any specific form of combination of hardware and software.

Finally, it should be noted that: the above examples are only specific embodiments of the present invention, and are not intended to limit the scope of the present invention, but it should be understood by those skilled in the art that the present invention is not limited thereto, and that the present invention is described in detail with reference to the foregoing examples: any person skilled in the art may modify or easily conceive of the technical solution described in the foregoing embodiments, or perform equivalent substitution of some of the technical features, while remaining within the technical scope of the present disclosure; such modifications, changes or substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention, and are intended to be included in the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. The signal lightweight automatic modulation recognition method based on deep learning is characterized by comprising the following steps of:

2. The deep learning-based signal lightweight automatic modulation recognition method according to claim 1, wherein obtaining a sample data set including a plurality of digital modulation schemes and a plurality of analog modulation schemes by simulating a real communication environment modulation information sequence comprises:

3. The deep learning-based signal lightweight automatic modulation recognition method according to claim 2, wherein a 01-bit information sequence is utilized to simulate a real communication environment modulation information sequence, and real communication environment influence parameters are added in the process of modulating and sampling the information sequence to obtain an IQ sampling sequence of a simulated modulation signal, wherein the real communication environment influence parameters include, but are not limited to: additive gaussian noise, multipath fading, sample rate offset, and center frequency offset.

4. The deep learning-based signal lightweight automatic modulation recognition method according to claim 2, wherein a sample data set is constructed by using N digital modulation signals and M analog modulation signals, and for information points of each modulation signal, K information points are continuously acquired at each time by taking a preset threshold number of information points as sampling intervals to form a signal sample; each modulation signal samples a preset number of samples, and the samples sampled by all the modulation signals are combined to construct a signal sample data set, wherein the preset number is set in units of ten thousands.

5. The deep learning-based signal lightweight automatic modulation recognition method according to claim 1, wherein the lightweight dense convolution long-short time memory network structure constructed based on the lightweight convolution network comprises: the method comprises the steps of inputting data subjected to standardization processing to a batch normalization processing unit, carrying out convolution and splicing operation on the standardized data, and extracting time sequence features of the spliced data by using a long-short-time memory network and classifying the time sequence features by activating a function.

6. The deep learning-based signal lightweight automatic modulation recognition method according to claim 5, wherein the feature extraction classification unit is formed by connecting two long-short-time memory network layers and a full-connection layer, the two long-short-time memory network layers adopt different numbers of hidden units, and the full-connection layer activation function adopts a Softmax function.

7. The deep learning-based signal lightweight automatic modulation recognition method according to claim 1, 5 or 6, wherein when training and optimizing a lightweight dense convolution long-short-time memory network structure by using a sample data set, setting an absolute cross entropy loss function as a target loss function in a training process, selecting an Adam optimizer to optimize the network, and setting the maximum iteration round or early shutdown to manufacture as an iteration termination condition of training and optimizing.

8. A deep learning-based signal lightweight automatic modulation recognition system, comprising: a sample construction module, a model training module and a target recognition module, wherein,

9. An electronic device comprising a memory and a processor, said processor and said memory completing communication with each other via a bus; the memory stores program instructions executable by the processor, the processor invoking the program instructions to perform the method steps of any of claims 1-7.

10. A computer-readable storage medium, characterized in that the computer-readable storage medium has stored therein a computer program which, when executed by a processor, implements the method steps of any of claims 1-7.