Interference signal recognition method based on Transformer network

Abstract

The present invention proposes a method for identifying interference signals based on a Transformer network, which mainly solves the technical problems that the existing methods are complex, difficult and incomplete in feature extraction of interference signals. The scheme includes: 1) simulating communication signal data and interference signal data, constructing single interference and composite interference data sets, and dividing and labeling them; 2) constructing a local feature extraction module; 3) constructing a global feature fusion module; 4) constructing an interference identification network based on the local feature extraction module and the global feature fusion module; 5) using the constructed data set to train and verify the interference signal identification network model to obtain the final interference signal identification network model; 7) using the model to accurately identify each type of interference signal. The present invention can realize the comprehensive extraction of global and local features of interference signals, as well as the full fusion of time domain and frequency domain features, effectively improving the recognition accuracy of interference signals.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless communications, and further relates to electronic countermeasures, and specifically is an interference signal recognition method based on a Transformer network, which can be used to recognize the types of various common interference signals in electronic countermeasures. Background Art

[0002] Modern information-based electronic warfare aims to compete for information resources, so its primary task is to prevent or prevent the enemy from disrupting, suppressing and intercepting its own information to the greatest extent possible, while ensuring that the enemy successfully collects, uses and controls various information. In this process, both sides will inevitably use various electronic interferences to damage or even destroy the enemy's communication system, making it impossible for the enemy to receive normal or useful information and thus unable to carry out operations, thus wasting opportunities.

[0003] With the rapid development of communication technology, the types of interference imposed in electronic countermeasures are increasing and the forms are becoming more and more complex. Therefore, we need to take accurate and efficient anti-interference measures to cope with complex battlefield environments. Communication anti-interference technology generally includes three aspects of research: interference identification, interference decision-making and interference suppression. Among them, interference identification is an important prerequisite for the ultimate success in suppressing interference. The interference signals in the actual communication confrontation environment are complex and diverse, and general interference suppression methods cannot effectively suppress all types of interference. Therefore, reliable interference identification technology is required as a prerequisite for interference suppression.

[0004] The traditional interference identification methods commonly used at present all need to extract various features of the interference signal first, and then identify it through support vector machines, decision trees or machine learning methods. The feature extraction process is complex and difficult, and the extracted features are incomplete, which easily loses a lot of useful information and makes it impossible to accurately identify. In recent years, the research work of deep learning in the field of interference identification has also made certain progress. In the literature [Li X, Fu D, Mou W, et al. An Identification Method of Navigation Signal Interference Type based on SqueezeNet Model [C]. 2020 IEEE 9th Joint International Information Technology and Artificial Intelligence Conference (ITAIC). IEEE, 2020, 9: 875-880], Li X and Fu D et al. proposed an interference type recognition algorithm based on the SqueezeNet model. The basic unit of the SqueezeNet network is a modular convolution, called the Fire module. In this paper, the Fire module is used as the basic module for stacking, supplemented by maximum pooling, average pooling and ReLu activation function to construct a SqueezeNet convolutional network. The input data is the two-dimensional time-frequency image of the interference signal, and linear swept frequency interference, single-tone interference, Gaussian pulse and BPSK communication signals are classified. The experimental results show that when the interference-to-noise ratio is 0 dB, the recognition accuracy of the algorithm for the four signals reaches 99.5%. In the literature [Xu Guojin. Research on Typical Communication Interference Signal Identification Technology [D]. Chengdu: University of Electronic Science and Technology, 2018], Xu Guojin modified the convolution kernel size and pooling size into one dimension based on the classic Inception convolutional neural network, and used the power-normalized I and Q signals and I + Q + amplitude spectrum as the network inputs to study the recognition performance of the one-dimensional Inception network for several typical interferences under different input conditions. The experimental results show that the interference recognition rate is higher when the I, Q complex baseband signals + amplitude spectrum are used as the network inputs. After that, the corresponding interference recognition situation when the interference-to-noise ratio distribution of the training set samples is different is further studied. The results show that the wider the distribution range of the interference-to-noise ratio of the training set samples, the higher the recognition rate of the one-dimensional Inception network for various types of interference.Through in-depth analysis of current research results in the field of interference identification, it can be found that although deep learning methods have general advantages in identifying interference signal types, current work in this area is generally based on CNN networks based on convolutional calculations, and uses simple classical network models. There is no consideration of improving the structure of the neural network model based on the communication characteristics of the interference signal itself.

[0005] Transformer is a deep neural network model that has received widespread attention in recent years. It was first applied in the field of natural language processing. Transformer completely abandons the common feature extraction algorithm and replaces it with a self-attention mechanism. It can learn long-range dependency information in the sequence and is a model that considers the global information of the sequence. In recent years, Transformer has achieved good results in various fields, so it has received widespread attention and has been greatly promoted, and various variants have been derived. However, although the Transformer network can directly capture the long-range dependency information between long-distance items, for a long sequence, the Transformer network calculates the correlation between the current position and any other position in the sequence in the same standard, and does not highlight the impact of local correlation at adjacent positions on the final network learning results. In addition, most of the related applications of the Transformer network are to solve problems in the field of natural language processing or computer vision, and research in the field of communications needs to be further developed. Summary of the invention

[0006] The purpose of the present invention is to propose a method for identifying interference signals based on a Transformer network in view of the deficiencies in the prior art, so as to solve the problem that the process of extracting features in the existing identification methods is complex and difficult, and the extracted features are incomplete, and a large amount of useful information is easily lost, resulting in the inability to accurately identify the problem. The present invention provides an end-to-end, lighter-weight interference signal identification method that covers local and global information, integrates time domain and frequency domain features, has higher recognition accuracy, and has a faster recognition rate through structural innovation of the functional modules in the classic Transformer network framework.

[0007] The idea of ​​implementing the present invention is: first, communication signal data and interference signal data are simulated to construct a data set for training, verifying and ultimately testing the performance of the neural network in the method; then, based on the classic Transformer network as the basic framework, two main functional modules, namely the local feature extraction module CNN-Transformer and the global feature fusion module Self-Cross-Attention, are first constructed, and an interference signal recognition network is built with the two functional modules as the main modules. Finally, the network is trained, verified and tested accordingly on the self-constructed data set to achieve effective recognition of interference signals.

[0008] The present invention achieves the above-mentioned purpose by the following specific steps:

[0009] (1) Construct a single interference data set, including a single interference time domain data set, a single interference amplitude spectrum data set, and a single interference phase spectrum data set:

[0010] (1.1) Use MATLAB to generate a quadrature phase shift keying (QPSK) intermediate frequency digital modulation signal as the basic communication signal;

[0011] (1.2) Use MATLAB to generate single-tone interference, multi-tone interference, linear frequency sweep interference, partial-band noise interference, and noise frequency modulation interference; in the interference-to-noise ratio range of [-10dB, 20dB], generate n sample data at every 2dB interval for each interference, and each sample has 1024 sampling points in the time domain, and superimpose them with Gaussian white noise to obtain interference signals;

[0012] (1.3) superimposing the basic communication signal generated in step 1.1 and the interference signal obtained in step 1.2 to generate a single interference time domain data set;

[0013] (1.4) Each sample in the single interference time domain data set is subjected to Fourier transformation to obtain a single interference amplitude spectrum and a single interference phase spectrum data, which together with the single interference time domain data constitute a single interference data set required for network training, verification and testing in the present invention;

[0014] (2) Constructing a composite interference data set, including a composite interference time domain data set, a composite interference amplitude spectrum data set, and a composite interference phase spectrum data set:

[0015] (2.1) Select multi-tone interference, linear frequency sweep interference and partial frequency band noise interference from the five interference patterns in step 1.2, superimpose them in pairs to form three composite interference patterns, and superimpose all of these three types of interference to form a fourth composite interference pattern; in the interference-to-noise ratio range of [-10dB, 20dB], each composite interference generates n sample data at every 2dB interval, and each sample has 1024 sampling points in the time domain, and superimposes them with Gaussian white noise to obtain composite interference signals;

[0016] (2.2) superimposing the basic communication signal generated in step 1.1 and the composite interference signal obtained in step 2.1 to generate a composite interference time domain data set;

[0017] (2.3) Each sample in the composite interference time domain data set is subjected to Fourier transformation to obtain composite interference amplitude spectrum and composite interference phase spectrum data, which together with the composite interference time domain data constitute the composite interference data set required for network training, verification and testing in the present invention;

[0018] (3) Dataset division:

[0019] Each n samples in each interference-to-noise ratio in the single interference dataset and the composite interference dataset are divided into training set, validation set and test set in a ratio of 6:2:2;

[0020] (4) Labeling of the dataset:

[0021] Each sample is labeled with the unique hot encoding vector of each type of interference signal as the corresponding label;

[0022] (5) Constructing a Transformer-based interference signal recognition network model:

[0023] (5.1) Construct a CNN-Transformer local feature extraction module to extract local adjacent features of interference signals:

[0024] The time domain data, amplitude spectrum data and phase spectrum data in a single interference data set are used as the input of the local feature extraction module CNN-Transformer. Three identical convolutional layers are used to scan the three interference signal sequences respectively to extract the local feature information between adjacent sampling points. Then, batch normalization and ReLU activation function processing are performed after each convolution operation. Finally, the three-way output of the CNN-Transformer local feature extraction module is obtained, and then the three-way output is processed by position encoding as the input of the Self-Cross-Attention global feature fusion module.

[0025] (5.2) Construct a Self-Cross-Attention global feature fusion module to fuse the time domain and frequency domain features of the interference signal:

[0026] (5.2.1) Extract global features within each domain through the self-attention mechanism:

[0027] The interference signal time domain data f after the CNN-Transformer local feature extraction module t Construct the query vector Q through linear mapping t, the key vector K t , value vector V t , calculate the self-attention distribution value f t1 :

[0028]

[0029] First, Q t Vector and K t The vectors are multiplied to calculate the correlation between the two, and then the result is divided by d K is the key vector K t The dimension of the vector V is then normalized by SoftMax to obtain the weight value we need. Finally, the weight value is combined with the vector V t Perform point multiplication and weighted sum to obtain the attention distribution value of the point; for the value vector V with larger dimension t , the result after the dot product will also become larger, and the SoftMax function will fall into a very small gradient, divided by That is to avoid this problem; similarly, the self-attention distribution value of the amplitude spectrum data and phase spectrum data of the interference signal is calculated to obtain f a1 and f p1 ;

[0030] (5.2.2) Fusion of global features in time and frequency domains through cross-attention mechanism:

[0031] The time domain data feature vector sequence f after the self-attention mechanism t1 After linear mapping, the query vector Q is constructed t1 , similarly, the amplitude spectrum feature vector sequence f a1 and phase spectrum characteristic vector sequence f p1 After linear mapping, each constructs its own key vector K a1 , K p1 Sum value vector V a1 、V p1 , then Q t1 , K a1 and V a1 Perform cross-attention calculation to obtain a cross-fused feature vector f ta , Q t1 , K p1 and V p1 Perform cross attention calculation to obtain the feature vector f tp ; Afterwards, f ta After linear mapping, the query vector Q is constructed ta ,f tp After linear mapping, the key vector K is constructed tp Sum value vector Vtp , calculate the cross attention distribution value, and finally get the feature vector f after secondary cross fusion tap ;

[0032] (5.3) Construct an interference signal recognition network model based on CNN-Transformer and Self-Cross-Attention modules:

[0033] Taking the encoder structure of the classic Transformer model as the basic framework, the local feature extraction module CNN-Transformer and the global feature fusion module Self-Cross-Attention as the basic functional modules, and the time domain data, amplitude spectrum data and phase spectrum data in a single interference data set as the network input, an interference signal recognition network is constructed. The specific structure is: CNN-Transformer local feature extraction module->Batch Normalization->ReLU activation function->Positional Embedding->Self-Cross-Attention global feature fusion module->Residual connection->Layer Normalization->Fully connected layer Feed Forward->Residual connection->Layer Normalization->One-dimensional deconvolution ConvTranspose1d layer->Linear layer->SoftMax normalization;

[0034] (6) Use the training set to train the interference signal recognition network model:

[0035] (6.1) Input the single interference time domain training set, amplitude spectrum training set and phase spectrum training set obtained in step 3 into the interference signal recognition network model constructed in step (5), use the Adam optimization function to iteratively update the parameters of the model, and use the CrossEntropyLoss cross entropy loss function to calculate the training error value of each iteration until the loss function converges, and obtain the interference recognition network model trained under the initial parameter setting;

[0036] (6.2) Adjust the parameters of the interference identification network model and fully train the models with different parameters until the loss function converges;

[0037] (6.3) Input the composite interference time domain training set, amplitude spectrum training set and phase spectrum training set generated in step 3 into the interference signal recognition network model constructed in step 5, and fully train the model under different parameter conditions;

[0038] (7) Use the validation set to validate the interference signal recognition network model under different parameter settings and determine the optimal parameters:

[0039] Input the single interference and compound interference validation sets into the interference signal recognition network trained under different parameters in step 6 respectively, and judge the output results of each sample. The percentage of the number of correctly judged samples of each type of interference to the total number of samples is used as the evaluation index, that is, the recognition accuracy. Compare the recognition accuracy of each type of interference signal under different network parameters, and take the parameters corresponding to the highest recognition accuracy as the optimal model parameters, and take the trained model corresponding to the optimal parameters as the final interference signal recognition network model;

[0040] (8) Interference signal identification:

[0041] The single interference and composite interference test sets are used as the input of the final interference signal recognition network model to obtain the recognition accuracy of various interference signals.

[0042] Compared with the prior art, the present invention has the following advantages

[0043] First, the local feature extraction module CNN-Transformer designed by the present invention utilizes the local connection characteristics of the convolutional neural network to extract the local feature information between adjacent sampling points of the interference signal, and realizes the adaptation with the Transformer network without breaking the basic framework of the Transformer network, thereby making up for the deficiency of the Transformer network focusing on learning the long-range dependency information in the sequence and ignoring the adjacent position information, and realizing the comprehensive extraction of local and global features of the interference signal;

[0044] Second, the global feature fusion module Self-Cross-Attention designed by the present invention fully integrates the time domain and frequency domain features of the interference signal using the cross-attention mechanism, so that higher-dimensional and deeper features of the signal can be learned;

[0045] Third, since the present invention uses the local feature extraction module CNN-Transformer and the global feature fusion module Self-Cross-Attention as the main modules to build an interference signal recognition network, thanks to the local feature extraction module's ability to capture features between adjacent sampling points and the global feature fusion module's ability to fuse the time domain and frequency domain features of interference signals, the interference signal recognition network can effectively improve the recognition accuracy of various interference signals;

[0046] Fourth, since the interference signal recognition network in the present invention abandons the decoder part in the classic Transformer model and takes the encoder as the basic framework, the number of parameters is greatly reduced, which makes the network's calculation amount less, thereby effectively shortening the recognition time of the interference signal, which is of great significance in the actual combat environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a flowchart of the implementation process of the method of the present invention;

[0048] Figure 2 It is a structural schematic diagram of the CNN-Transformer local feature extraction module of the present invention;

[0049] Figure 3 Schematic diagram of the structure of the Self-Cross-Attention global feature fusion module of the present invention;

[0050] Figure 4 This is a schematic diagram of the interference signal identification network structure of the present invention;

[0051] Figure 5 It is a curve diagram of the recognition rate of various types of single interferences by the method of the present invention on a single interference data set;

[0052] Figure 6 It is a recognition rate curve diagram of various types of composite interferences by the method of the present invention on the composite interference data set;

[0053] Figure 7 This is a comparison chart of the recognition accuracy of the method of the present invention and the one-dimensional CNN network for each type of single interference;

[0054] Figure 8 This is a comparison chart of the recognition accuracy of the method of the present invention and the one-dimensional CNN network for each type of composite interference. DETAILED DESCRIPTION

[0055] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments:

[0056] Embodiment 1:

[0057] Reference Figure 1 , a method for interfering signal recognition based on Transformer network is implemented in the following steps:

[0058] Step 1: construct a single interference data set, including a single interference time domain data set, a single interference amplitude spectrum data set, and a single interference phase spectrum data set:

[0059] (1.1) Use MATLAB to generate a quadrature phase shift keying (QPSK) intermediate frequency digital modulation signal as the basic communication signal;

[0060] (1.2) Use MATLAB to generate single-tone interference, multi-tone interference, linear frequency sweep interference, partial-band noise interference, and noise frequency modulation interference; in the interference-to-noise ratio range of [-10dB, 20dB], each type of interference generates n sample data every 2dB, and each sample has 1024 sampling points in the time domain, which are superimposed with Gaussian white noise to obtain interference signals. The center frequency, period, bandwidth and other parameters of each type of interference signal are randomly distributed; the five types of interference generate a total of 5×16×n sample data under 16 interference-to-noise ratios;

[0061] (1.3) superimposing the basic communication signal generated in step 1.1 and the interference signal obtained in step 1.2 to generate a single interference time domain data set;

[0062] (1.4) Each sample in the single interference time domain data set is subjected to Fourier transformation to obtain a single interference amplitude spectrum and a single interference phase spectrum data, which together with the single interference time domain data constitute a single interference data set required for network training, verification and testing in the present invention;

[0063] Step 2: construct a composite interference data set, including a composite interference time domain data set, a composite interference amplitude spectrum data set, and a composite interference phase spectrum data set:

[0064] (2.1) Select multi-tone interference, linear frequency sweep interference and partial frequency band noise interference from the five interference patterns in step 1.2, superimpose them in pairs to form three composite interference patterns, and superimpose all of these three types of interference to form a fourth composite interference pattern; in the interference-to-noise ratio range of [-10dB, 20dB], each composite interference generates n sample data at every 2dB interval, and each sample has 1024 sampling points in the time domain, and superimposes them with Gaussian white noise to obtain composite interference signals;

[0065] (2.2) superimposing the basic communication signal generated in step 1.1 and the composite interference signal obtained in step 2.1 to generate a composite interference time domain data set;

[0066] (2.3) Each sample in the composite interference time domain data set is subjected to Fourier transformation to obtain composite interference amplitude spectrum and composite interference phase spectrum data, which together with the composite interference time domain data constitute the composite interference data set required for network training, verification and testing in the present invention;

[0067] Step 3, data set division:

[0068] Each n samples in each interference-to-noise ratio in the single interference dataset and the composite interference dataset are divided into training set, validation set and test set in a ratio of 6:2:2;

[0069] Step 4: Label the dataset:

[0070] Each sample is labeled with the unique hot encoding vector of each type of interference signal as the corresponding label;

[0071] Step 5: Build a Transformer-based interference signal recognition network model:

[0072] (5.1) Construct a CNN-Transformer local feature extraction module to extract local adjacent features of interference signals:

[0073] The time domain data, amplitude spectrum data and phase spectrum data in a single interference data set are used as the input of the local feature extraction module CNN-Transformer. Three identical convolutional layers are used to scan the three interference signal sequences respectively to extract the local feature information between adjacent sampling points. Then, batch normalization and ReLU activation function processing are performed after each convolution operation. Finally, the three-way output of the CNN-Transformer local feature extraction module is obtained, and then the three-way output is processed by position encoding as the input of the Self-Cross-Attention global feature fusion module.

[0074] The specific parameters of the convolutional layer in the CNN-Transformer local feature extraction module are as follows:

[0075] The number of input channels of the one-dimensional convolution is set to 1, the number of output channels of the one-dimensional convolution, that is, the number of convolution kernels, is set to 128, the convolution kernel size is set to 1×3, the sliding step size is set to 1, and the input matrix size before entering the one-dimensional convolution is (64,1,1024), where 64 is the batch size, and 1024 is the number of sampling points of each data sample, that is, the sample length; 128 1×3 convolution kernels are used to scan the 1024 sampling points of each one-dimensional data to extract the local features between adjacent sampling points. Finally, 128-dimensional local feature information is extracted for every 3 adjacent points. Therefore, the output matrix size of the one-dimensional convolution module is (64,128,1024); the output of the one-dimensional convolution is used as the input of the position encoding, and after the position information is loaded into the sequence, it is used as the input of the Self-Cross-Attention global feature fusion module;

[0076] (5.2) Construct a Self-Cross-Attention global feature fusion module to fuse the time domain and frequency domain features of the interference signal:

[0077] (5.2.1) Extract global features within each domain through the self-attention mechanism:

[0078] The interference signal time domain data f after the CNN-Transformer local feature extraction module tConstruct the query vector Q through linear mapping t , the key vector K t , value vector V t , calculate the self-attention distribution value f t1 :

[0079]

[0080] First, Q t Vector and K t The vectors are multiplied to calculate the correlation between the two, and then the result is divided by d K is the key vector K t The dimension of the vector V is then normalized by SoftMax to obtain the weight value we need. Finally, the weight value is combined with the vector V t Perform point multiplication and weighted sum to obtain the attention distribution value of the point; for the value vector V with larger dimension t , the result after the dot product will also become larger, and the SoftMax function will fall into a very small gradient, divided by That is to avoid this problem; similarly, the self-attention distribution value of the amplitude spectrum data and phase spectrum data of the interference signal is calculated to obtain f a1 and f p1 ;

[0081] (5.2.2) Fusion of global features in time and frequency domains through cross-attention mechanism:

[0082] The time domain data feature vector sequence f after the self-attention mechanism t1 After linear mapping, the query vector Q is constructed t1 , similarly, the amplitude spectrum feature vector sequence f a1 and phase spectrum characteristic vector sequence f p1 After linear mapping, each constructs its own key vector K a1 , K p1 Sum value vector V a1 、V p1 , then Q t1 , K a1 and V a1 Perform cross-attention calculation to obtain a cross-fused feature vector f ta , Q t1 , K p1 and V p1 Perform cross attention calculation to obtain the feature vector f tp ; Afterwards, f ta After linear mapping, the query vector Q is constructed ta ,f tp After linear mapping, the key vector K is constructedtp Sum value vector V tp , calculate the cross attention distribution value, and finally get the feature vector f after secondary cross fusion tap ;

[0083] (5.3) Construct an interference signal recognition network model based on CNN-Transformer and Self-Cross-Attention modules:

[0084] Taking the encoder structure of the classic Transformer model as the basic framework, the local feature extraction module CNN-Transformer and the global feature fusion module Self-Cross-Attention as the basic functional modules, and the time domain data, amplitude spectrum data and phase spectrum data in a single interference data set as the network input, an interference signal recognition network is constructed. The specific structure is: CNN-Transformer local feature extraction module->Batch Normalization->ReLU activation function->Positional Embedding->Self-Cross-Attention global feature fusion module->Residual connection->Layer Normalization->Fully connected Feed Forward layer->Residual connection->Layer Normalization->One-dimensional deconvolution ConvTranspose1d layer->Linear layer->SoftMax normalization;

[0085] In the interference signal recognition network constructed based on CNN-Transformer and Self-Cross-Attention modules, each sublayer is connected through a residual module; the input channel and output channel parameter settings of the final deconvolution layer are opposite to the one-dimensional convolution settings in the CNN-Transformer module, that is, the number of input channels is 128, the number of output channels is 1, and the convolution kernel size and step length remain unchanged, which are 1×3 and 1 respectively; after that, a linear layer is passed to obtain an output result with a dimension of the number of types of interference signals to be identified, and then the values ​​are normalized by SoftMax, the value size is judged, and finally the identification result of the interference signal type is obtained;

[0086] Step 6: Use the training set to train the interference signal recognition network model:

[0087] (6.1) Input the single interference time domain training set, amplitude spectrum training set and phase spectrum training set obtained in step 3 into the interference signal recognition network model constructed in step (5), use the Adam optimization function to iteratively update the parameters of the model, and use the CrossEntropyLoss cross entropy loss function to calculate the training error value of each iteration until the loss function converges, and obtain the interference recognition network model trained under the initial parameter setting;

[0088] (6.2) Adjust the parameters of the interference identification network model and fully train the models with different parameters until the loss function converges;

[0089] (6.3) Input the composite interference time domain training set, amplitude spectrum training set and phase spectrum training set generated in step 3 into the interference signal recognition network model constructed in step 5, and fully train the model under different parameter conditions;

[0090] Step 7: Use the validation set to validate the interference signal recognition network model under different parameter settings to determine the optimal parameters:

[0091] Input the single interference and compound interference validation sets into the interference signal recognition network trained under different parameters in step 6 respectively, and judge the output results of each sample. The percentage of the number of correctly judged samples of each type of interference to the total number of samples is used as the evaluation index, that is, the recognition accuracy. Compare the recognition accuracy of each type of interference signal under different network parameters, and take the parameters corresponding to the highest recognition accuracy as the optimal model parameters, and take the trained model corresponding to the optimal parameters as the final interference signal recognition network model;

[0092] Step 8, interference signal identification:

[0093] The single interference and composite interference test sets are used as the input of the final interference signal recognition network model to obtain the recognition accuracy of various interference signals.

[0094] Embodiment 2:

[0095] Reference Figure 1 The overall implementation steps of this embodiment are the same as those of the first embodiment, and the specific implementation process of the present invention is further described by giving specific settings for the parameters:

[0096] Step A: Construct a single interference dataset, including a time domain dataset, an amplitude spectrum dataset, and a phase spectrum dataset:

[0097] (a.1) Use MATLAB to generate a quadrature phase shift keying (QPSK) intermediate frequency digital modulation signal with the following parameters: carrier frequency is 2 MHz, sampling frequency is 8 MHz, roll-off coefficient of shaping filter is 0.8, and number of sampling points is 1024;

[0098] (a.2) Five common interference signals are generated by MATLAB, namely single-tone interference, multi-tone interference, linear frequency sweep interference, partial-band noise interference and noise frequency modulation interference, and superimposed with Gaussian white noise of a certain power. In the interference-to-noise ratio range of [-10dB, 20dB], 1000 sample data are generated for each interference at every 2dB interval. Each sample has 1024 sampling points in the time domain. The center frequency, period, bandwidth and other parameters of each simulated interference signal are randomly distributed. Specifically, the main simulation parameters of various interference signals are shown in Table 1:

[0099] Table 1 Interference signal simulation parameters

[0100]

[0101] (a.3) superimposing the QPSK signal simulated and generated in step (a.1) with the interference signals under different interference-to-noise ratio conditions simulated and generated in step (a.2) to generate a single interference time domain data set;

[0102] (a.4) performing Fourier transformation on each sample in the single interference time domain data set generated in (a.3) to obtain amplitude spectrum and phase spectrum data, which together with the time domain data constitute the single interference data set required for network training, verification and testing in the present invention;

[0103] Step B: Construct a composite interference data set, which also includes a time domain data set, an amplitude spectrum data set, and a phase spectrum data set:

[0104] (b.1) Multi-tone interference, linear frequency sweep interference and partial frequency band noise interference are selected from the five interference patterns simulated in (a.2), and three composite interference patterns are formed by superimposing two of them. In addition, all three types of interference are superimposed to form a fourth composite interference pattern. Similarly, they are superimposed with Gaussian white noise of a certain power. Within the interference-to-noise ratio range of [-10dB, 20dB], each composite interference generates 1000 sample data at every 2dB interval. Each sample has 1024 sampling points in the time domain. They are superimposed with the QPSK signal simulated in (a.1) to generate a composite interference time domain data set.

[0105] (b.2) performing Fourier transformation on each sample in the composite interference time domain data set generated in (b.1) to obtain amplitude spectrum and phase spectrum data, which together with the time domain data constitute the composite interference data set required for network training, verification and testing in the present invention;

[0106] Step C: Dataset division:

[0107] Divide every 1000 samples in each interference-to-noise ratio condition in the single interference and composite interference data sets constructed in step A and step B into a training set, a validation set, and a test set in a ratio of 6:2:2;

[0108] Step D: Labeling of the dataset:

[0109] Each sample is labeled with the unique-hot encoding vector of each type of interference signal. The labels corresponding to the five interference patterns in the single interference data set are [1,0,0,0,0], [0,1,0,0,0], [0,0,1,0,0], [0,0,0,1,0], [0,0,0,0,1], and [0,0,0,0,1] respectively. The labels corresponding to the four types of composite interference in the composite interference data set are [1,1,0], [1,0,1], [0,1,1], and [1,1,1] respectively.

[0110] Step E: Constructing a Transformer-based interference signal recognition network model:

[0111] (e.1) Construct a CNN-Transformer local feature extraction module to extract local adjacent features of the interference signal:

[0112] The input of the constructed local feature extraction module CNN-Transformer is a one-dimensional interference signal. The one-dimensional convolution kernel is used to scan the interference data sequence to extract the local feature information between adjacent sampling points. At the same time, it replaces the word embedding module in the classic Transformer model to expand the data dimension and achieve adaptation with the Transformer network. The output of the one-dimensional convolution is then used as the input of the Transformer bottom encoder after passing through the position encoding module. At the same time, in order to avoid gradient dissipation and ensure the consistent distribution of the intermediate layer results, batch normalization and ReLU activation function processing are performed after the convolution operation. The specific parameters of the convolution layer in the CNN network are as follows: the number of input channels is set to 1, representing an input communication interference data sample. For the classic Transformer model, the vector dimension of word embedding is generally 128 or 256, which is the optimal dimension size of the Transformer model. Therefore, the number of output channels of the one-dimensional convolution, that is, the number of convolution kernels, is set to 128 to facilitate the subsequent fusion with the position encoding module. The remaining parameters are set as follows: the convolution kernel size is set to 1×3, the sliding step size is set to 1, and the input matrix size before entering the one-dimensional convolution is (64,1,1024), where 64 is the batch size Batch Size, 1024 is the number of sampling points of each data sample, that is, the sample length. 128 1×3 convolution kernels are used to scan the 1024 sampling points of each one-dimensional data to extract the local features between adjacent sampling points. Finally, 128-dimensional local feature information is extracted from every 3 adjacent points. Therefore, the output matrix size of the one-dimensional convolution module is (64, 128, 1024). After one-dimensional convolution, the local features of the interference signal are extracted, and the vector dimension is also adapted to the Transformer encoder. The output of the one-dimensional convolution is used as the input of the position encoding. After the position information is loaded on the sequence, the Q, K, and V matrices are obtained through linear transformation, and the attention distribution value is calculated to extract the global information of the interference signal.

[0113] (e.2) Construct a Self-Cross-Attention global feature fusion module to fuse the time domain and frequency domain features of the interference signal:

[0114] The input of the constructed global feature fusion module Self-Cross-Attention is the feature vector sequence f of the time domain data, amplitude spectrum data and phase spectrum data of the interference signal after passing through the CNN-Transformer local feature extraction module. t , f a and f p , in the Self-Cross-Attention module, f t , f a and fp First, the global feature information in each domain is extracted through the self-attention mechanism to obtain f t1 , f a1 and f p1 , and then f t1 and f a1 、f t1 and f p1 Perform a cross-attention operation to obtain a cross-fused feature vector f ta and f tp Finally, f ta and f tp Perform another cross-attention operation to obtain the feature vector f after the second cross-fusion tap , the vector fully integrates the time domain and frequency domain characteristics of the interference signal, and its specific steps include:

[0115] (e.2.1) Extract global features within each domain through the self-attention mechanism:

[0116] The time domain data f of the interference signal after the CNN-Transformer local feature extraction module t After linear mapping, the query vector Q is constructed t , the key vector K t , value vector V t , calculate the self-attention distribution value f t1 :

[0117]

[0118] First, Q t Vector and K t The vectors are multiplied to calculate the correlation between the two, and then the result is divided by d K is the key vector K t The dimension of the vector V is then normalized by SoftMax to obtain the weight value we need. Finally, the weight value is combined with the vector V t Perform point multiplication and weighted sum to get the attention score value of the point. For the value vector V with larger dimension t , the result after the dot product will also become larger, and the SoftMax function will fall into a very small gradient, divided by That is to avoid this problem; similarly, the self-attention distribution value of the amplitude spectrum data and phase spectrum data of the interference signal is calculated to obtain f a1 and f p1 ;

[0119] (e.2.2) Fusion of global features in time and frequency domains through cross-attention mechanism:

[0120] The time domain data feature vector sequence f after the self-attention mechanism t1 After linear mapping, the query vector Q is constructed t1 , similarly, the amplitude spectrum feature vector sequence f a1 and phase spectrum characteristic vector sequence f p1 After linear mapping, each constructs its own key vector K a1 , K p1 Sum value vector V a1 、V p1 , then Q t1 , K a1 and V a1 Perform cross-attention calculation to obtain a cross-fused feature vector f ta , Q t1 , K p1 and V p1 Perform cross attention calculation to obtain the feature vector f tp ; Afterwards, f ta After linear mapping, the query vector Q is constructed ta ,f tp After linear mapping, the key vector K is constructed tp Sum value vector V tp , calculate the cross attention distribution value, and finally get the feature vector f after secondary cross fusion tap ;

[0121] (e.3) Constructing an interference signal recognition network based on CNN-Transformer and Self-Cross-Attention modules: The interference signal recognition network constructed by the present invention is based on the encoder structure of the classic Transformer model, with the local feature extraction module CNN-Transformer and the global feature fusion module Self-Cross-Attention as the basic functional modules, and the time domain data, amplitude spectrum data and phase spectrum data of the interference signal as the network input. The input data of the three channels first pass through the local feature extraction module CNN-Transformer to extract the local feature information between the adjacent sampling points, and then enter the global feature fusion module respectively, wherein the global feature information in the domain is first extracted by the self-attention mechanism, and then the global features on different domains are fully fused through the cross-attention mechanism, so as to realize the comprehensive extraction of local features and global features, and the full fusion of time domain features and frequency domain features; The specific structure of the interference signal recognition network constructed by the present invention is: CNN-Transformer local feature extraction module->Batch Normalization->ReLU activation function->Positional encoding Embedding->Self-Cross-Attention global feature fusion module->Residual connection->Layer Normalization->Fully connected layerFeed Forward->Residual connection->Layer Normalization->One-dimensional deconvolutionConvTranspose1d layer->Linear layerLinear->SoftMax normalization, where each sub-layer is connected through a residual module to slow down network degradation and avoid gradient dissipation. The input channel and output channel parameter settings of the final deconvolution layer are opposite to those of the one-dimensional convolution in the CNN-Transformer module, that is, the number of input channels is 128, the number of output channels is 1, and the convolution kernel size and step length remain unchanged, which are 1×3 and 1 respectively. After that, a linear layer is passed to obtain an output result with a dimension of the number of types of interference signals to be identified. After the values ​​are normalized by SoftMax, the value size is judged, and finally the recognition result of the type of interference signal is obtained.

[0122] Step F: Interference signal recognition network training:

[0123] (f.1) Input the single interference time domain training set, amplitude spectrum training set and phase spectrum training set generated by simulation in step A into the interference signal recognition network based on Transformer constructed in step E, use the Adam optimization function to iteratively update the parameters of the model, and use the CrossEntropyLoss cross entropy loss function to calculate the training error value of each iteration until the loss function converges, thereby obtaining a trained interference recognition network model;

[0124] (f.2) Adjust the parameters of the model and fully train the model under different parameter conditions;

[0125] (f.3) inputting the composite interference time domain training set, amplitude spectrum training set and phase spectrum training set generated by simulation in step B into the interference signal recognition network based on Transformer constructed in step E, and fully training the model under different parameter conditions;

[0126] Step G: Interference signal identification network parameter verification:

[0127] The single interference and compound interference verification sets are respectively input into the interference signal recognition network trained under different parameters in step F, and the output results of each sample are judged. The percentage of the number of correctly judged samples of each type of interference to the total number of samples is used as the evaluation index, which is called the recognition accuracy. The recognition accuracy of each type of interference signal under different network parameters is compared to determine the optimal model parameters. The optimal network parameters finally verified are shown in Table 2:

[0128] Table 2 Optimal network parameters

[0129] Parameter name Parameter size Number of training iterations 30 Training set capacity 48000 Training batch size 64 Learning Rate 0.0001 Loss Function CrossEntropyLoss Optimizing functions Adam Function Convolutional layer output dimension 128 Number of encoder layers 6 Number of heads in multi-head attention mechanism 8 Q, K, V vector dimensions 16

[0130] Step H: Interference signal identification network test:

[0131] Load the trained model corresponding to the optimal network parameters verified in step G, use the single interference and composite interference test sets simulated and generated in steps A and B as inputs to the interference identification network, and obtain the identification accuracy of various interference signals;

[0132] Step I: Interference signal identification network test effect comparison:

[0133] A four-layer one-dimensional convolutional neural network is constructed, and corresponding training, verification and testing are performed on the constructed data set. The test results of the method of the present invention and the one-dimensional convolutional neural network are compared to prove the superior performance of the method of the present invention. The constructed convolutional network is composed of four Conv+Pooling modules stacked together, and the activation function uses the ReLU function to appropriately alleviate the gradient vanishing problem, and the convergence speed is fast, easy to implement, and also improves the learning efficiency. Pooling uses the maximum pooling method. The specific parameter sizes of each layer are shown in Table 3:

[0134] Table 3 CNN network parameters

[0135]

[0136] Then, through a fully connected layer, the learned features are classified and summarized, and finally the type of interference signal is identified after SoftMax normalization.

[0137] The effect of the present invention is further described below in conjunction with simulation experiments.

[0138] 1. Experimental conditions:

[0139] The hardware platform of the simulation experiment of the present invention is: H3C Uniserver R5300 G5 server, the processor is Intel Xeon Gold 5218 dual CPU, the main frequency is 2.3GHz, the memory is 512GB, three Nvidia GeForce RTX2080Ti graphics cards, each graphics card has 12GB memory;

[0140] The software platform of the simulation experiment of the present invention is: Windows Server 2019 operating system, the integrated development platform is Pycharm 2021, and the deep learning framework used is Pytorch.

[0141] 2. Experimental content:

[0142] Using the self-constructed data set, the one-dimensional CNN network and the interference signal recognition network of the present invention are trained respectively and their respective recognition performances are tested to compare the performance differences between the one-dimensional CNN network and the interference recognition method of the present invention.

[0143] 3. Evaluation indicators:

[0144] The output result of each sample is judged, and the percentage of the number of correctly judged samples of various types of interference to the total number of samples is used as the evaluation index, which is called recognition accuracy.

[0145] 4. Experimental results and analysis:

[0146] Figure 5 The recognition accuracy curve of each type of single interference under different interference-to-noise ratios for the single interference data set constructed by the method of the present invention is shown in the figure. As can be seen from the figure, when the interference-to-noise ratio is greater than -2dB, the recognition accuracy of the five types of interference signals is close to 100%. Under the low interference-to-noise ratio of -10dB, the recognition accuracy of each type of interference is above 70%, and the recognition rate curve of each type of interference shows a steady upward trend with the increase of the interference-to-noise ratio, which is relatively stable as a whole without obvious fluctuations.

[0147] Figure 6 The following is a curve diagram of the recognition accuracy of each composite interference under different interference-to-noise ratios on the composite interference data set constructed by the method of the present invention. It can be seen from the figure that when the interference-to-noise ratio is greater than 0dB, the recognition accuracy of the four composite interferences is close to 100%, among which, for the case where all three interferences are superimposed, the recognition accuracy is the lowest. This is because the features are complicated due to the superposition of different interferences, so it is not easy to accurately distinguish the correct category of the interference signal from the various features that are mixed together.

[0148] Figure 7 The figure is a comparison chart of the recognition accuracy of the method of the present invention and the one-dimensional CNN network for five single interferences.

[0149] Figure 8 This is a comparison chart of the recognition accuracy of the method of the present invention and the one-dimensional CNN network for four types of composite interferences.

[0150] from Figure 7 and Figure 8 It can be seen that in the single interference and compound interference data sets constructed by themselves, under various interference-to-noise ratio conditions, the method of the present invention shows a higher recognition accuracy than the one-dimensional CNN network. Especially for the recognition of four types of compound interference, the advantages of the method of the present invention are more obvious. Not only is the recognition accuracy significantly higher than that of the one-dimensional CNN network, but the overall curve rises steadily with the increase of the interference-to-noise ratio. The recognition rate curve of the one-dimensional CNN network for compound interference shows obvious fluctuations and repetitions, which is relatively unstable. Furthermore, whether it is for the recognition of single interference or compound interference, as the interference-to-noise ratio decreases, the recognition accuracy of the one-dimensional CNN network decreases significantly, which is much greater than the decrease corresponding to the method of the present invention.

[0151] The above experimental results prove the correctness and effectiveness of the method proposed in the present invention.

[0152] Parts of the present invention that are not described in detail belong to common knowledge among those skilled in the art.

[0153] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. It is obvious that for professionals in this field, after understanding the content and principles of the present invention, they may make various modifications and changes in form and details without departing from the principles and structures of the present invention. However, these modifications and changes based on the ideas of the present invention are still within the scope of protection of the claims of the present invention.

Claims

1. A method for identifying interference signals based on a Transformer network, characterized in that: The Transformer network is applied to the field of signal processing. The encoder structure of the classic Transformer network is used as the basic framework. The interference recognition network is constructed in combination with the characteristics of the interference signal. Single interference and composite interference data sets are constructed for network training, verification and testing. The specific steps include: (1) Construct a single interference data set, including a single interference time domain data set, a single interference amplitude spectrum data set, and a single interference phase spectrum data set: (1.1) Use MATLAB to generate a quadrature phase shift keying (QPSK) intermediate frequency digital modulation signal as the basic communication signal; (1.2) Use MATLAB to generate single-tone interference, multi-tone interference, linear frequency sweep interference, partial-band noise interference, and noise frequency modulation interference; in the interference-to-noise ratio range of [-10dB, 20dB], generate n sample data at every 2dB interval for each interference, and each sample has 1024 sampling points in the time domain, and superimpose them with Gaussian white noise to obtain interference signals; (1.3) superimposing the basic communication signal generated in step (1.1) and the interference signal obtained in step (1.2) to generate a single interference time domain data set; (1.4) Each sample in the single interference time domain data set is subjected to Fourier transform to obtain a single interference amplitude spectrum and a single interference phase spectrum data, which together with the single interference time domain data constitute a single interference data set required for network training, verification and testing; (2) Constructing a composite interference data set, including a composite interference time domain data set, a composite interference amplitude spectrum data set, and a composite interference phase spectrum data set: (2.1) Select multi-tone interference, linear frequency sweep interference and partial frequency band noise interference from the five interference patterns in step (1.2), superimpose them in pairs to form three composite interference patterns, and superimpose all three types of interference patterns to form a fourth composite interference pattern; in the interference-to-noise ratio interval of [-10dB, 20dB], each composite interference generates n sample data at every 2dB interval, and each sample has 1024 sampling points in the time domain, and superimposes them with Gaussian white noise to obtain composite interference signals; (2.2) superimposing the basic communication signal generated in step (1.1) and the composite interference signal obtained in step (2.1) to generate a composite interference time domain data set; (2.3) Each sample in the composite interference time domain data set is subjected to Fourier transform to obtain composite interference amplitude spectrum and composite interference phase spectrum data, which together with the composite interference time domain data constitute the composite interference data set required for network training, verification and testing; (3) Dataset division: Each n samples in each interference-to-noise ratio in the single interference dataset and the composite interference dataset are divided into training set, validation set and test set in a ratio of 6:2:2; (4) Labeling of datasets: Each sample is labeled with the unique hot encoding vector of each type of interference signal as the corresponding label; (5) Constructing a Transformer-based interference signal recognition network model: (5.1) Construct a CNN-Transformer local feature extraction module to extract local adjacent features of interference signals: The time domain data, amplitude spectrum data and phase spectrum data in a single interference data set are used as the input of the local feature extraction module CNN-Transformer. Three identical convolutional layers are used to scan the three interference signal sequences respectively to extract the local feature information between adjacent sampling points. Then, batch normalization and ReLU activation function processing are performed after each convolution operation. Finally, the three-way output of the CNN-Transformer local feature extraction module is obtained, and then the three-way output is processed by position encoding as the input of the Self-Cross-Attention global feature fusion module. (5.2) Construct a Self-Cross-Attention global feature fusion module to fuse the time domain and frequency domain features of the interference signal: (5.2.1) Extract global features within each domain through the self-attention mechanism: The interference signal time domain data f after the CNN-Transformer local feature extraction module t Construct the query vector Q through linear mapping t , the key vector K t , value vector V t , calculate the self-attention distribution value f t1 : Calculate the self-attention distribution value of the amplitude spectrum data and phase spectrum data of the interference signal to obtain f a1 and f p1 ; (5.2.2) Fusion of global features in time and frequency domains through cross-attention mechanism: The time domain data feature vector sequence f after the self-attention mechanism t1 After linear mapping, the query vector Q is constructed t1 , similarly, the amplitude spectrum feature vector sequence f a1 and phase spectrum characteristic vector sequence f p1 After linear mapping, each constructs its own key vector K a1 , K p1 Sum value vector V a1 、V p1 , then Q t1 , K a1 and V a1 Perform cross-attention calculation to obtain a cross-fused feature vector f ta , Q t1 , K p1 and V p1 Perform cross attention calculation to obtain the feature vector f tp ; Afterwards, f ta After linear mapping, the query vector Q is constructed ta ,f tp After linear mapping, the key vector K is constructed tp Sum value vector V tp , calculate the cross attention distribution value, and finally get the feature vector f after secondary cross fusion tap ; (5.3) Construct an interference signal recognition network model based on CNN-Transformer and Self-Cross-Attention modules: Taking the encoder structure of the classic Transformer model as the basic framework, the local feature extraction module CNN-Transformer and the global feature fusion module Self-Cross-Attention as the basic functional modules, and the time domain data, amplitude spectrum data and phase spectrum data in a single interference data set as the network input, an interference signal recognition network is constructed. The specific structure is: CNN-Transformer local feature extraction module->Batch Normalization->ReLU activation function->Positional Embedding->Self-Cross-Attention global feature fusion module->Residual connection->Layer Normalization->Fully connected layer Feed Forward->Residual connection->Layer Normalization->One-dimensional deconvolution ConvTranspose1d layer->Linear layer->SoftMax normalization; (6) Use the training set to train the interference signal recognition network model: (6.1) Input the single interference time domain training set, amplitude spectrum training set and phase spectrum training set obtained in step (3) into the interference signal recognition network model constructed in step (5), use the Adam optimization function to iteratively update the parameters of the model, and use the CrossEntropyLoss cross entropy loss function to calculate the training error value of each iteration until the loss function converges, thereby obtaining the interference recognition network model trained under the initial parameter settings; (6.2) Adjust the parameters of the interference identification network model and fully train the models with different parameters until the loss function converges; (6.3) Inputting the composite interference time domain training set, amplitude spectrum training set and phase spectrum training set generated in step (3) into the interference signal recognition network model constructed in step (5), and fully training the model under different parameter conditions; (7) Use the validation set to validate the interference signal recognition network model under different parameter settings and determine the optimal parameters: The single interference and composite interference validation sets are respectively input into the interference signal recognition network trained under different parameters in step (6), and the output results of each sample are judged. The percentage of the number of correctly judged samples of each type of interference to the total number of samples is used as the evaluation index, i.e., the recognition accuracy. The recognition accuracy of each type of interference signal under different network parameters is compared, and the parameters corresponding to the highest recognition accuracy are used as the optimal model parameters, and the trained model corresponding to the optimal parameters is used as the final interference signal recognition network model; (8) Interference signal identification: The single interference and composite interference test sets are used as the input of the final interference signal recognition network model to obtain the recognition accuracy of various interference signals.

2. The method according to claim 1, characterized in that: In step (1.1), MATLAB is used to generate a quadrature phase shift keying (QPSK) intermediate frequency digital modulation signal, and its parameters are set as follows: the carrier frequency is 2 MHz, the sampling frequency is 8 MHz, the roll-off coefficient of the shaping filter is 0.8, and the number of sampling points is 1024.

3. The method according to claim 1, characterized in that: In step (1.2), MATLAB is used to generate single-tone interference, multi-tone interference, linear swept frequency interference, partial-band noise interference and noise frequency modulation interference. The center frequency, period, bandwidth and other parameters of each interference signal are randomly distributed; the five types of interference generate a total of 5×16×n sample data under 16 interference-to-noise ratios.

4. The method according to claim 1, characterized in that: In step (4), each sample is labeled as follows: the labels corresponding to the five interference patterns in the single interference data set are [1,0,0,0,0], [0,1,0,0,0], [0,0,1,0,0], [0,0,0,1,0], [0,0,0,0,1]; the labels corresponding to the four types of composite interference in the composite interference data set are [1,1,0], [1,0,1], [0,1,1], [1,1,1].

5. The method according to claim 1, characterized in that: The specific parameters of the convolutional layer in the CNN-Transformer local feature extraction module in step (5.1) are as follows: The number of input channels of the one-dimensional convolution is set to 1, the number of output channels of the one-dimensional convolution, that is, the number of convolution kernels, is set to 128, the convolution kernel size is set to 1×3, the sliding step size is set to 1, and the input matrix size before entering the one-dimensional convolution is (64,1,1024), where 64 is the batch size Batch Size, and 1024 is the number of sampling points of each data sample, that is, the sample length; 128 1×3 convolution kernels are used to scan the 1024 sampling points of each one-dimensional data to extract local features between adjacent sampling points. Finally, 128-dimensional local feature information is extracted for every 3 adjacent points. Therefore, the output matrix size of the one-dimensional convolution module is (64,128,1024); the output of the one-dimensional convolution is used as the input of the position encoding, and after the position information is loaded into the sequence, it is used as the input of the Self-Cross-Attention global feature fusion module.

6. The method according to claim 1, characterized in that: In the interference signal recognition network based on CNN-Transformer and Self-Cross-Attention modules constructed in step (5.3), each sublayer is connected by a residual module; the input channel and output channel parameter settings of the final deconvolution layer are opposite to the settings of the one-dimensional convolution in the CNN-Transformer module, that is, the number of input channels is 128, the number of output channels is 1, and the convolution kernel size and step length remain unchanged, which are 1×3 and 1 respectively; after that, after passing through a linear layer, an output result with a dimension of the number of types of interference signals to be identified is obtained, and then after the numerical value is normalized by SoftMax, the numerical size is judged, and finally the recognition result of the type of interference signal is obtained.

Images