Modulation coding joint recognition method and system based on multi-channel attention network

Abstract

The present application relates to the technical field of signal modulation and coding joint identification, and particularly relates to a modulation and coding joint identification method and system based on a multi-channel attention network, which encodes a modulation information sequence by simulating a communication environment with a randomly generated radio information sequence, generates a digital modulation signal and an analog modulation signal, and assembles signal sample data; constructs a multi-channel convolutional gated deep attention network model for modulation and coding joint identification, and trains and optimizes the multi-channel convolutional gated deep attention network model using the signal sample data; inputs a signal to be identified into the trained and optimized multi-channel convolutional gated deep attention network model, and obtains the modulation and coding category of the signal to be identified using the trained and optimized multi-channel convolutional gated deep attention network model. The present application can improve the efficiency of modulation and coding joint identification, can realize the task of signal modulation and coding joint automatic identification, and is convenient for deployment and implementation.

Description

Technical Field

[0001] This invention relates to the field of signal modulation and coding joint recognition technology, and in particular to a modulation and coding joint recognition method and system based on a multi-channel attention network. Background Technology

[0002] With the development of communication technology, Variable Code Modulation (AVM) and Adaptive Code Modulation (ACM) technologies have emerged to improve communication capacity and are widely used in communications. Adaptive Code Modulation can flexibly select the coding and modulation method according to changes in channel transmission conditions, greatly improving spectrum utilization. In interactive point-to-point applications, using ACM technology can increase satellite communication capacity by 100% to 200%. With the continuous strengthening of modern informatization and digitalization, satellite communication networks are vulnerable to third-party interception, interference, and attacks. To enhance the security of satellite communication networks, unintentional interference can be used to identify the modulation and coding methods of communication signals, thereby obtaining relevant signal parameters. This can provide decision-making guidance for improving satellite communication management, strengthening inter-system coordination, and optimizing anti-interference performance.

[0003] Currently, modulation-coding joint identification of signals can be divided into two categories: traditional hierarchical identification algorithms and deep learning-based modulation-coding joint identification algorithms. Traditional algorithms primarily employ a hierarchical identification approach, first identifying the modulation of the signal, then demodulating it, and finally identifying the coding type and estimating the parameters of the demodulated information stream to achieve joint identification of modulation and coding. With hierarchical identification, the coding recognition performance depends to some extent on the modulation identification performance and the demodulation algorithm error; a decrease in modulation identification performance or a large demodulation algorithm error inevitably leads to a decrease in coding identification performance. Deep learning-based modulation-coding joint identification mostly uses intermediate frequency (IF) signal waveforms as input to deep networks. However, IF signals are less conducive to feature extraction by deep networks compared to baseband quadrature-in-phase (I / Q) signals. This is because the joint constraints of modulation and coding in a signal are inherently difficult to extract. The carrier frequency in an IF signal occupies a large portion of the signal's energy, increasing the difficulty of extracting these constraints, a problem not present in baseband I / Q signals. Meanwhile, current deep network structures used for modulation-coding joint recognition lack specificity and have insufficient ability to extract joint modulation-coding features, resulting in low recognition rates and robustness of the algorithms. Summary of the Invention

[0004] To address this, the present invention provides a modulation-coding joint recognition method and system based on a multi-channel attention network, which solves the problems of unreasonable signal input form and poor deep network structure design in existing modulation-coding joint recognition methods, resulting in unsatisfactory recognition rate and robustness.

[0005] According to the design scheme provided by this invention, a modulation-coding joint recognition method based on a multi-channel attention network is provided, comprising:

[0006] The communication environment is simulated by randomly generated radio information sequences, which are then encoded and modulated to generate digital and analog modulated signals.

[0007] Signal sample data is constructed using digital modulation signals and analog modulation signals, and the signal sample data is divided into training sample set, verification sample set and test sample set according to a preset ratio;

[0008] A multi-channel convolutional gated deep attention network model for modulation-coding joint recognition was constructed, and the model was trained and optimized using training, validation, and test sample sets.

[0009] The signal to be identified is input into the trained and optimized multi-channel convolutional gated deep attention network model, and the modulation and coding category of the signal to be identified is obtained by using the trained and optimized multi-channel convolutional gated deep attention network model.

[0010] As part of the modulation-coding joint identification method based on a multi-channel attention network of this invention, further, it utilizes randomly generated radio information sequences to simulate the communication environment, encodes modulation information sequences, and generates digital modulation signals and analog modulation signals, comprising:

[0011] First, a 0-1 bit radio information sequence is randomly generated using a random function;

[0012] Next, in a simulated communication environment, noise interference is added during the coding, modulation, and sampling of the radio information sequence to obtain the IQ sampling sequence of the modulated signal. The noise interference includes additive Gaussian noise and the influence of the Rayleigh fading channel model.

[0013] Then, for the IQ sampling sequence, multiple digital modulation signals are generated based on the digital modulation signal code rate, which are combinations of various channel codes and modulation methods, and an analog modulation signal without channel coding is generated.

[0014] As a modulation-coding joint identification method based on multi-channel attention network of the present invention, the digital modulation signal further includes 28 digital modulation signals formed by combining four channel codes (LDPC code, RS code, TCM code and TPC code) and seven modulation methods (2FSK, GMSK, BPSK, QPSK, 8PSK, 8APSK and 16APSK); the analog modulation signal includes AM and FM signals without channel coding.

[0015] As a modulation-coded joint recognition method based on multi-channel attention network of the present invention, further, signal sample data is constructed using digital modulation signals and analog modulation signals, and the signal sample data is divided into training sample set, verification sample set and test sample set according to a preset ratio. This includes: for each type of modulation signal information point, m information points are continuously collected at a sampling interval of n information points to form a signal sample; M signal samples are collected for each type of modulation signal, and a signal sample dataset is constructed based on the signal samples of all modulation signals; sample data from each type of modulation-coded signal are extracted from the signal sample dataset according to a preset ratio to form the training sample set, verification sample set and test sample set, where m, n and M are preset thresholds.

[0016] As a modulation-coding joint recognition method based on multi-channel attention network of the present invention, the constructed multi-channel convolutional gated deep attention network model includes: a multi-channel convolution module for extracting feature maps from input data through multi-channel convolution operations; a dimension-up convolution module for performing dimension-up convolution on the extracted feature maps to extract target shapes and details in the feature maps; a gated recurrent classification module for capturing deep connection dimension information and classifying and outputting modulation categories; and a squeeze excitation module for adaptively recalibrating channel weights by explicitly modeling the interdependence between channels using channel weight vectors.

[0017] As a modulation-coding joint recognition method based on multi-channel attention network of the present invention, the multi-channel convolution module extracts feature maps from the input data through multi-channel convolution operation, including: using three convolutional layers to convolve the I-channel signal, Q-channel signal and I / Q signal in the input data respectively, and concatenating the outputs of the three convolutional layers according to the input dimension to obtain the feature map of the input data.

[0018] As a modulation-coding joint recognition method based on multi-channel attention network of the present invention, the upscaling convolution module performs upscaling convolution on the extracted feature map, including: upsampling the input low-resolution feature map sequentially using three concatenated convolution units to output a high-resolution feature map, wherein the convolution unit includes a one-dimensional convolutional layer, a random deactivation layer, a batch normalization layer and a ReLU activation function.

[0019] As a modulation-coding joint recognition method based on multi-channel attention network of the present invention, the gated recurrent classification module captures the deep connection dimension information and classifies and outputs the modulation category, including: using a gated recurrent unit (GRU) to combine the current input and the hidden state passed by the previous node to output the output of the current hidden node and pass the hidden state of the next node, so as to capture the feature map dimension information of deep connection through gating state update; and using a fully connected output layer to classify the feature map to obtain the corresponding signal modulation category.

[0020] As a modulation-coding joint recognition method based on a multi-channel attention network in this invention, the process by which the squeezing excitation module recalibrates the sequence feature channel weight vector based on the degree of feature sharing in the classification task in the input feature vector to generate a new feature vector is further represented as follows: in, For the new feature vector The i-th element, s i • The i-th element X in the channel weight vector S of the input feature vector X i The channel weights are S = σ(W2δ(W1Z)), where Z is the channel statistics vector generated by global average pooling of the output feature vector of the first layer on the network, δ is the ReLU function, σ is the Sigmoid activation function, W1 is the weight parameter of the first fully connected layer in the squeeze activation module, and W2 is the weight parameter of the second fully connected layer in the squeeze activation module.

[0021] Furthermore, the present invention also provides a modulation-coding joint recognition system based on a multi-channel attention network, comprising: a data simulation module, a data processing module, a model building module, and a target recognition module, wherein,

[0022] The data simulation module is used to simulate the communication environment by encoding modulation information sequences and generating digital and analog modulation signals using randomly generated radio information sequences.

[0023] The data processing module is used to construct signal sample data using digital modulation signals and analog modulation signals, and to divide the signal sample data into training sample set, verification sample set and test sample set according to a preset ratio.

[0024] The model building module is used to build a multi-channel convolutional gated deep attention network model for modulation-coding joint recognition and to train and optimize the multi-channel convolutional gated deep attention network model using training sample sets, validation sample sets, and test sample sets;

[0025] The target recognition module is used to input the signal to be recognized into the trained and optimized multi-channel convolutional gated deep attention network model, and to obtain the modulation and coding category of the signal to be recognized using the trained and optimized multi-channel convolutional gated deep attention network model.

[0026] The beneficial effects of this invention are:

[0027] This invention addresses the issues of unreasonable signal input formats and poor deep network structure design in existing modulation-coding joint recognition methods, which lead to unsatisfactory recognition rates and robustness. By designing a multi-channel convolutional gated deep attention network model and training and optimizing it using simulated digital signals and analog signal sample data, the trained network model structure can be used to identify the modulation type of the target signal. This enables automatic joint recognition of signal modulation and coding, and is easy to deploy and implement. Attached image description:

[0028] Figure 1 This is a schematic diagram of the modulation-coding joint recognition principle based on a multi-channel attention network in the embodiment;

[0029] Figure 2 This is a schematic diagram of the multi-channel convolutional gated deep attention network model structure in the embodiment;

[0030] Figure 3 This is a schematic diagram of the construction process of the multi-channel convolutional gated deep attention network model in the embodiment;

[0031] Figure 4 This is a schematic diagram illustrating the change in validation loss during the network training process in the embodiment;

[0032] Figure 5 This is a schematic diagram of the recognition rate results of modulation and coding joint recognition in the embodiment. Detailed implementation method:

[0033] To make the objectives, technical solutions, and advantages of this invention clearer and more understandable, the invention will be further described in detail below with reference to the accompanying drawings and technical solutions.

[0034] To address the issues of unreasonable signal input formats and poor deep network structure design leading to unsatisfactory recognition rates and robustness in existing modulation-code joint recognition methods, this invention provides a modulation-code joint recognition method based on a multi-channel attention network, comprising:

[0035] S101. Using randomly generated radio information sequences to simulate the communication environment, the information sequence is encoded and modulated to generate digital and analog modulated signals.

[0036] Specifically, it can be designed to include the following content:

[0037] First, a 0-1 bit radio information sequence is randomly generated using a random function;

[0038] Next, in a simulated communication environment, noise interference is added during the coding, modulation, and sampling of the radio information sequence to obtain the IQ sampling sequence of the modulated signal. The noise interference includes additive Gaussian noise and the influence of the Rayleigh fading channel model.

[0039] Then, for the IQ sampling sequence, multiple digital modulation signals are generated based on the digital modulation signal code rate, which are combinations of various channel codes and modulation methods, and an analog modulation signal without channel coding is generated.

[0040] The information sequence of a randomly generated radio signal is produced using a random function to generate a 0-1 bit sequence, ensuring the randomness of the information content and eliminating the influence of signal content on the joint identification of signal modulation and coding. The information sequence is encoded and modulated in a simulated real communication environment, and orthogonal in-phase (IQ) sequences are obtained through sampling. Additive Gaussian noise and the influence of Rayleigh fading channel model are incorporated into the encoding, modulation, and sampling processes to simulate the real environment. During the simulation, parameters can be set as follows: modulation rate 2 MBaud / s, sampling rate 10 MHz, and signal-to-noise ratio (SNR) ranging from -8 dB to 20 dB in 2 dB increments. Finally, the IQ sampling sequence of the modulated signal is obtained, i.e., 5 information points per symbol.

[0041] The analog-generated digital modulation signal contains 28 combined modulation and coding signals, employing four channel codes (LDPC code with a code rate of 1 / 2, RS code with a code rate of 56 / 64, TCM code with a code rate of 2 / 3, and TPC code with a code rate of 676 / 1024) and seven modulation schemes (2FSK, GMSK, BPSK, QPSK, 8PSK, 8APSK, and 16APSK). The analog modulation signals are uncoded AM and FM signals.

[0042] S102. Construct signal sample data using digital modulation signals and analog modulation signals, and divide the signal sample data into training sample set, verification sample set and test sample set according to a preset ratio.

[0043] Specifically, it can be designed to include the following: for each type of modulation signal information point, m information points are continuously collected at a sampling interval of n information points to form a signal sample; M signal samples are collected for each modulation signal, and a signal sample dataset is constructed based on the signal samples of all modulation signals; sample data from each type of modulation-coded signal are extracted from the signal sample dataset according to a preset ratio to form a training sample set, a verification sample set, and a test sample set, where m, n, and M are preset thresholds.

[0044] In practical applications, for each type of modulation and coding signal, 10 information points can be sampled at intervals of 10 information points, with 1024 information points collected consecutively each time to form a signal sample. For each type of modulation and coding signal, signals are collected at 2dB intervals between -8dB and 20dB, with 1000 samples collected at each signal-to-noise ratio, totaling 15000 samples for each type of modulation and coding signal. All the signals are then combined into a signal sample set. The total number of samples in the signal sample set is 450,000. From the generated sample set, 60% of each type of modulation and coding signal is extracted to form a training sample set, 20% is extracted from the remaining 40% to form a validation sample set, and the last 20% of the entire sample set is used as the test sample set. For each signal-to-noise ratio under each modulation and coding scheme, 1000 signal samples were first randomly selected and added to the training set. Then, 20% of the remaining 40% were randomly selected and added to the validation set. Finally, the remaining 20% ​​were added to the test set. This increased the randomness of the sample data and improved the training quality of the network model.

[0045] S103. Construct a multi-channel convolutional gated deep attention network model for modulation-coding joint recognition and train and optimize the multi-channel convolutional gated deep attention network model using training sample sets, validation sample sets, and test sample sets.

[0046] Specifically, the constructed multi-channel convolutional gated deep attention network model structure is as follows: Figure 2 As shown, it includes: a multi-channel convolution module for extracting feature maps from input data through multi-channel convolution operations; a dimension-increasing convolution module for performing dimension-increasing convolution on the extracted feature maps to extract the target shape and details in the feature maps; a gated recurrent classification module for capturing deep connection dimension information and classifying and outputting modulation categories; and a squeeze excitation module for adaptively recalibrating channel weights by explicitly modeling the interdependencies between channels using channel weight vectors.

[0047] The multi-channel convolution module extracts feature maps from the input data through multi-channel convolution operations, including: using three convolutional layers to convolve the I-channel signal, Q-channel signal, and I / Q signal in the input data respectively, and concatenating the outputs of the three convolutional layers according to the input dimension to obtain the feature map of the input data.

[0048] Each of the three one-dimensional convolutional layers (ConvI, ConvQ, ConvI / Q) can be followed by a ReLU activation function. The connection structure of the multi-channel convolutional module can be described as follows: I-channel signal of normalized I / Q signal → ConvI → Concatenate; Q-channel signal of normalized I / Q signal → ConvQ → Concatenate; normalized I / Q signal → ConvI / Q → Concatenate; Concatenate → SE block1.

[0049] The upscaling convolution module performs upscaling convolution on the extracted feature map, which includes: upsampling the input low-resolution feature map sequentially using three concatenated convolution units to output a high-resolution feature map. The convolution unit includes a one-dimensional convolutional layer, a random deactivation layer, a batch normalization layer, and a ReLU activation function.

[0050] The connection structure of the dimension-upgrading convolution module can be described as: SE block1 → convolution unit 1 → convolution unit 2 → convolution unit 3 → SE block2. Among them, SE block1 and SE block2 are two excitation modules.

[0051] The gated recurrent classification module captures the dimensional information of deep connections and classifies and outputs the modulation category. It includes: using a gated recurrent unit (GRU) to combine the current input and the hidden state passed from the previous node to output the output of the current hidden node and pass the hidden state of the next node, so as to capture the dimensional information of the feature map of deep connections through gating state updates; and using a fully connected output layer to classify the feature map to obtain the corresponding signal modulation category.

[0052] The gated recurrent classification module can be composed of a gated recurrent unit (GRU1), a random deactivation layer, a gated recurrent unit (GRU2), and a fully connected output layer. The activation function of the fully connected output layer is Softmax.

[0053] The squeeze excitation module applies the feature vector X∈R output by the previous layer of the neural network. L×C By global average pooling F sq (·) can generate channel statistical vectors Z∈R 1×C Where L is the length of the feature vector, C represents the number of channels in the feature vector, and the i-th element z of Z is... i The feature vector is calculated using the following formula, where L is the length of the feature vector and C represents the number of channels in the feature vector.

[0054]

[0055] Z can be transformed by the specific change F shown in the following formula. exThe weight vector S for the X channels of the sequence feature is generated using (·, W), where δ refers to the ReLU function and σ is the Sigmoid activation function. Each value in the weight vector represents the result of the SE block learning the importance of each channel of X for the classification task; the larger the value, the more beneficial the channel is for the classification task.

[0056] S = F ex (Z,W)=σ(g(Z,W))=σ(W2δ(W1Z))

[0057] After obtaining the weight vector S of the channels of the sequence feature X, S can be used to recalibrate the channel weights of X through multiplication, generating a new feature vector. The i-th element The following formula is used to calculate S, where S = [s1, s2, ..., s C Compared to X, The weights of each channel are recalibrated based on their contribution to the classification task, so that... It is more targeted to the final classification task.

[0058]

[0059] In practical applications, the parameters of each module in the network model can be designed as follows:

[0060] The one-dimensional convolutional layer ConvI has 8 convolutional kernels and a kernel size of 3.

[0061] The one-dimensional convolutional layer ConvQ has 8 convolutional kernels and a kernel size of 3.

[0062] The one-dimensional convolutional layer ConvI / Q has 16 convolutional kernels and a kernel size of 3.

[0063] In SE block1, r = 4 is set.

[0064] In convolutional unit 1, the number of convolutional kernels in the one-dimensional convolutional layer is 160, the kernel size is 7, and the convolution stride is 4.

[0065] In convolutional unit 2, the number of convolutional kernels in the one-dimensional convolutional layer is 200, the kernel size is 7, and the convolution stride is 4.

[0066] The number of convolution kernels in the one-dimensional convolutional layer of convolutional unit 3 is 256, the kernel size is 5, and the convolution stride is 2.

[0067] In SE block2, r = 4 is set.

[0068] GRU1 has 256 hidden units and returns the entire output sequence.

[0069] GRU1 has 256 hidden units and returns the last output of the output sequence.

[0070] The number of convolutional kernels in the fully connected layer is 30, corresponding to the number of output categories.

[0071] When training a multi-channel convolutional gating deep attention network model using the training set, the Adam optimizer can be selected to optimize the network. The initial learning rate is set to 0.001, with 512 samples per batch and a maximum of 100 training epochs for the entire training dataset. The model is validated using a validation set after each training epoch, with the validation loss used as a reference. Training is stopped when the validation loss does not decrease after 20 training epochs.

[0072] The order of all samples in the training dataset can be shuffled, and the training and validation samples can be input into a multi-channel convolutionally gated deep attention network model to train the network. The training process is completed when the maximum number of training iterations is reached or the early stopping mechanism is met, resulting in a trained multi-channel convolutionally gated deep attention network model. Test sample sets can then be input into the trained multi-channel convolutionally gated deep attention network to obtain recognition results, which can be compared with the true categories to statistically evaluate the network model's recognition accuracy.

[0073] S104. Input the signal to be identified into the trained and optimized sequence fusion convolutional network model, and use the trained and optimized sequence fusion convolutional network model to obtain the modulation and coding category of the signal to be identified.

[0074] Furthermore, based on the above method, this embodiment of the invention also provides a modulation-coding joint recognition system based on a multi-channel attention network, comprising: a data simulation module, a data processing module, a model building module, and a target recognition module, wherein,

[0075] The data simulation module is used to simulate the communication environment by encoding modulation information sequences and generating digital and analog modulation signals using randomly generated radio information sequences.

[0076] The data processing module is used to construct signal sample data using digital modulation signals and analog modulation signals, and to divide the signal sample data into training sample set, verification sample set and test sample set according to a preset ratio.

[0077] The model building module is used to build a multi-channel convolutional gated deep attention network model for modulation-coding joint recognition and to train and optimize the multi-channel convolutional gated deep attention network model using training sample sets, validation sample sets, and test sample sets;

[0078] The target recognition module is used to input the signal to be recognized into the trained and optimized multi-channel convolutional gated deep attention network model, and to obtain the modulation and coding category of the signal to be recognized using the trained and optimized multi-channel convolutional gated deep attention network model.

[0079] To verify the effectiveness of this solution, the following explanation is based on experimental data:

[0080] Simulation experiments were implemented on NVIDIA Quadro RTX 6000 and Keras 2.6.0 Tensorflow-GPU 2.4.0 platforms, completing the simulation of modulation signal generation and multi-channel convolutional gating deep attention network training in the embodiment of this case. Figure 3 The experiment was completed by steps (1) to (5) shown, and the trend of the verification loss during the training process of the multi-channel convolutional gated deep attention network and the recognition rate of the modulation coding joint recognition method based on the multi-channel convolutional gated deep attention network were obtained.

[0081] Figure 4 The graph illustrates the change in validation loss during network training. As shown, the validation loss decreases, converges, and stabilizes, indicating that the training effect of the simulation experiment gradually improves with increasing training iterations. Figure 5 The simulation results show that the recognition rate gradually increases and stabilizes as the signal-to-noise ratio increases, reaching a maximum of 93.7%. The model's inference speed was also recorded as 0.024 milliseconds per sample.

[0082] The experimental data above further verifies that the proposed solution can complete the automatic modulation recognition task. The solution is effective and feasible, and can be applied in practical scenarios such as signal reconnaissance and / or defense and electronic countermeasures.

[0083] Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention.

[0084] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0085] The units and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations are not considered to be beyond the scope of this invention.

[0086] Those skilled in the art will understand that all or part of the steps in the above methods can be implemented by a program instructing related hardware, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk. Optionally, all or part of the steps in the above embodiments can also be implemented using one or more integrated circuits. Accordingly, each module / unit in the above embodiments can be implemented in hardware or as a software functional module. This invention is not limited to any particular combination of hardware and software.

[0087] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A modulation-coding joint recognition method based on a multi-channel attention network, characterized in that, Include: The communication environment is simulated by randomly generated radio information sequences, which are then encoded and modulated to generate digital and analog modulated signals. Signal sample data is constructed using digital modulation signals and analog modulation signals, and the signal sample data is divided into training sample set, verification sample set and test sample set according to a preset ratio; A multi-channel convolutional gated deep attention network model for modulation-coding joint recognition is constructed and trained and optimized using training, validation, and test sample sets. The constructed multi-channel convolutional gated deep attention network model includes: a multi-channel convolution module for extracting feature maps from input data through multi-channel convolution operations; a dimensionality-increasing convolution module for performing dimensionality-increasing convolutions on the extracted feature maps to extract target shapes and details from the feature maps; a gated recurrent classification module for capturing deep connection dimension information and classifying modulation categories; and a squeeze excitation module for adaptively recalibrating channel weights by explicitly modeling the interdependencies between channels using channel weight vectors. The multi-channel convolution module extracts feature maps from input data through multi-channel convolution operations, including: using three convolutional layers to convolve the I-channel signal, Q-channel signal, and I / Q signal in the input data, respectively, and... The outputs of the three convolutional layers are concatenated along the input dimension to obtain the feature map of the input data. The upsampling convolution module performs upsampling on the extracted feature map, including: sequentially upsampling the input low-resolution feature map using three concatenated convolutional units to output a high-resolution feature map, wherein the convolutional unit includes a one-dimensional convolutional layer, a random deactivation layer, a batch normalization layer, and a ReLU activation function. The gated recurrent classification module captures the dimensional information of deep connections and classifies the modulation category, including: using a gated recurrent unit (GRU) to combine the current input and the hidden state passed from the previous node to output the output of the current hidden node and pass the hidden state of the next node, so as to capture the dimensional information of the feature map of deep connections through gating state updates; and using a fully connected output layer to classify the feature map to obtain the corresponding signal modulation category. The squeeze excitation module uses the channel weight vector to explicitly model the interdependence between channels and adaptively recalibrate the channel weights, which is represented as: , For the new feature vector The One element, Input feature vector Channel weight vector The Middle element The channel weights, and , This is a channel statistics vector generated by global average pooling of the output feature vector of the first layer of the network. For ReLU function, It is the Sigmoid activation function. These are the weight parameters of the first fully connected layer in the squeeze excitation module. These are the weight parameters of the second fully connected layer in the extrusion excitation module; The signal to be identified is input into the trained and optimized multi-channel convolutional gated deep attention network model, and the modulation and coding category of the signal to be identified is obtained by using the trained and optimized multi-channel convolutional gated deep attention network model.

2. The modulation-coding joint recognition method based on a multi-channel attention network according to claim 1, characterized in that, Using randomly generated radio information sequences to simulate a communication environment, a modulated information sequence is encoded and generated to produce digital and analog modulated signals, including: First, a 0-1 bit radio information sequence is randomly generated using a random function; Next, in a simulated communication environment, noise interference is added during the coding, modulation, and sampling of the radio information sequence to obtain the IQ sampling sequence of the modulated signal. The noise interference includes additive Gaussian noise and the influence of the Rayleigh fading channel model. Then, for the IQ sampling sequence, multiple digital modulation signals are generated based on the digital modulation signal code rate, which are combinations of various channel codes and modulation methods, and an analog modulation signal without channel coding is generated.

3. The modulation-coding joint recognition method based on a multi-channel attention network according to claim 1 or 2, characterized in that, The digital modulation signal includes 28 digital modulation signals formed by combining four channel codes (LDPC, RS, TCM, and TPC) and seven modulation methods (2FSK, GMSK, BPSK, QPSK, 8PSK, 8APSK, and 16APSK); the analog modulation signal includes AM and FM signals without channel coding.

4. The modulation-coding joint recognition method based on a multi-channel attention network according to claim 1, characterized in that, Signal sample data is constructed using digital and analog modulation signals, and the signal sample data is divided into training sample set, verification sample set and test sample set according to a preset ratio. It includes: for each modulation signal information point, m information points are continuously collected at a sampling interval of n information points to form a signal sample; M signal samples are collected for each modulation signal, and a signal sample dataset is constructed based on the signal samples of all modulation signals. Sample data from each type of modulation and coding signal are extracted from the signal sample dataset according to a preset ratio to form a training sample set, a validation sample set, and a test sample set, where m, n, and M are preset thresholds.

5. A modulation-coding joint recognition system based on a multi-channel attention network, characterized in that, It includes: a data simulation module, a data processing module, a model building module, and a target recognition module. The data simulation module is used to simulate the communication environment by encoding modulation information sequences and generating digital and analog modulation signals using randomly generated radio information sequences. The data processing module is used to construct signal sample data using digital modulation signals and analog modulation signals, and to divide the signal sample data into training sample set, verification sample set and test sample set according to a preset ratio. The model building module is used to construct a multi-channel convolutional gated deep attention network model for modulation-coding joint recognition and to train and optimize the multi-channel convolutional gated deep attention network model using training sample sets, validation sample sets, and test sample sets. The constructed multi-channel convolutional gated deep attention network model includes: a multi-channel convolution module for extracting feature maps from input data through multi-channel convolution operations; a dimensionality-increasing convolution module for performing dimensionality-increasing convolutions on the extracted feature maps to extract target shapes and details from the feature maps; a gated recurrent classification module for capturing deep connection dimension information and classifying and outputting modulation categories; and a squeeze excitation module for adaptively recalibrating channel weights by explicitly modeling the interdependencies between channels using channel weight vectors. The multi-channel convolution module extracts feature maps from input data through multi-channel convolution operations, including: using three convolutional layers to convolve the I-channel signal, Q-channel signal, and I / Q signal in the input data, respectively. The input data is processed by concatenating the outputs of three convolutional layers according to the input dimension to obtain a feature map. The upsampling convolution module performs upsampling on the extracted feature map, including: sequentially upsampling the input low-resolution feature map using three concatenated convolutional units to output a high-resolution feature map, wherein the convolutional unit includes a one-dimensional convolutional layer, a random deactivation layer, a batch normalization layer, and a ReLU activation function. The gated recurrent classification module captures the dimensional information of deep connections and classifies the modulation category, including: using a gated recurrent unit (GRU) to combine the current input and the hidden state passed from the previous node to output the output of the current hidden node and pass the hidden state of the next node, so as to capture the dimensional information of the feature map of deep connections through gating state updates; and using a fully connected output layer to classify the feature map to obtain the corresponding signal modulation category. The squeeze excitation module uses the channel weight vector to explicitly model the interdependence between channels and adaptively recalibrate the channel weights, which is represented as: , For the new feature vector The One element, Input feature vector Channel weight vector The Middle element The channel weights, and , This is a channel statistics vector generated by global average pooling of the output feature vector of the first layer of the network. For ReLU function, It is the Sigmoid activation function. These are the weight parameters of the first fully connected layer in the squeeze excitation module. These are the weight parameters of the second fully connected layer in the extrusion excitation module; The target recognition module is used to input the signal to be recognized into the trained and optimized multi-channel convolutional gated deep attention network model, and to obtain the modulation and coding category of the signal to be recognized using the trained and optimized multi-channel convolutional gated deep attention network model.

Images