Communication signal classification method based on local feature window fusion ViT network

Abstract

This invention discloses a communication signal classification method based on local feature window fusion ViT network. The method first obtains signal samples under different signal-to-noise ratios (SNRs), performs carrier synchronization and symbol synchronization processing, and transforms them into baseband sequences with additive white Gaussian noise. Next, after normalization preprocessing, the baseband sequence is converted into a constellation diagram. Then, the constellation diagram is divided into N patches, and position encoding is performed on each patch to obtain the input sequence for the Transformer encoder. Finally, the global and local feature components obtained from the two branches of the encoder are combined to construct a loss function, yielding the signal classification result. This invention solves the problems of classification difficulties caused by noise in low SNR environments and poor classification performance of high-order signals, and it can extract features more clearly, improving classification accuracy.

Description

Technical Field

[0001] This invention belongs to the field of deep learning and relates to a communication signal classification method based on local feature window fusion ViT network. Background Technology

[0002] The establishment of modern communication systems requires ensuring the effective transmission of information, which is achieved by modulating baseband signals to improve communication quality. Before acquiring information content, it is necessary to first determine the modulation pattern of the received signal. Therefore, automatic modulation classification technology becomes a crucial step before signal demodulation and has significant research value. Since there are significant differences in the phase and amplitude information characteristics between baseband signals of different modulation types, these differences can be visually represented by constellation diagrams. Therefore, constellation diagram features can be extracted using deep learning networks to classify signals of different modulation types.

[0003] In their paper "Modulation Classification Based on SignalConstellation Diagrams and Deep Learning," Peng S et al. used CNN as the network model and conducted an in-depth analysis of this classification method. However, the AlexNet model they chose only had 8 layers, resulting in weak fitting ability and poor performance in classifying high-order signals. Huang G et al., in their paper "Modulation classification of MQAM signals based on gradient color constellation and deep learning," proposed a method to classify multi-level orthogonal amplitude modulation (MQAM) signals by extracting constellation diagram features using an 18-layer Residual Network (ResNet). This method deepened the network layers and avoided the gradient vanishing problem through residual learning. However, ResNet suffers from the same problem as AlexNet: extracting constellation point features ignores global contextual information. In low signal-to-noise ratio (SNR) environments, local constellation point information is severely affected by noise, blurring the phase and amplitude information of the signal. The network lacks attention to the global distribution of constellation points, leading to a significant performance degradation. Traditional Vision Transformer (ViT) models have the advantage of fully utilizing global contextual information, but they lack the local receptive field characteristics of Convolutional Neural Networks (CNNs). Using traditional ViT networks for constellation map classification will result in ignoring the correlation between local constellation points. Summary of the Invention

[0004] The purpose of this invention is to improve the signal classification capability of signal sets containing {Binary Phase Shift Keying (BPSK), Quadrature Phase Shift Keying (QPSK), 8-Phase Shift Keying (8PSK), 16-Amplitude Phase Shift Keying (16APSK), 32-Amplitude Phase Shift Keying (32APSK), 16-Quadrature Amplitude Modulation (16QAM), 32-Quadrature Amplitude Modulation (32QAM), and 64-Quadrature Amplitude Modulation (64QAM)}, and proposes a communication signal classification method based on local feature window fusion of ViT networks. This method incorporates a local feature window fusion module to enhance the correlation between local features while retaining the advantage of the original ViT network in fully utilizing global information. By leveraging both global and local information, it achieves accurate differentiation between different modulation types. Under varying signal-to-noise ratios, compared to AlexNet, ResNet-18, ResNet-50, and ViT networks, the proposed network model significantly improves classification accuracy for BPSK, QPSK, 8PSK, 16APSK, 32APSK, 16QAM, 32QAM, and 64QAM.

[0005] The technical solution adopted by this invention to solve its technical problem includes the following steps:

[0006] Step 1: Obtain communication signal samples with different signal-to-noise ratios through real receivers and simulations, and after carrier synchronization and symbol synchronization processing, transform them into baseband sequences with additive white Gaussian noise.

[0007] Step 2: After the baseband sequence is normalized, the preprocessed baseband sequence is converted into a constellation diagram.

[0008] Step 3: Divide the constellation diagram into N patches, perform position encoding on each patch, and obtain the input sequence of the Transformer encoder.

[0009] Step 4: Input the input sequence into a Transformer encoder. It first passes through an l-layer Transformer network. Then, the output features of the l-th Transformer network layer are input into two branches: one branch passes through a Global Feature Extraction (GFE) module and a Transformer network layer; the other branch passes through a Local Feature Extraction (LFE) module and a Transformer network layer. These two branches output the global feature components g, respectively. class With local feature components

[0010] Step 5: Calculate the obtained global feature components g class With local feature components Input the classifier, construct the loss function, and obtain the classification result of the communication signal.

[0011] The specific implementation of the signal model received in step 1 is as follows:

[0012] Communication signal samples with different signal-to-noise ratios were obtained through real receivers and simulations. Carrier synchronization and symbol synchronization were performed, and the signal samples were converted into baseband sequences with additive white Gaussian noise. The signal sample set includes 8 modulation types: BPSK, QPSK, 8PSK, 16APSK, 32APSK, 16QAM, 32QAM, and 64QAM. The general signal model s(n) is as follows:

[0013]

[0014] Where a(n) represents the amplitude of the nth symbol; f c Let f be the center frequency of the signal. Assuming the signal has undergone carrier synchronization processing, we consider f to be... c Approaching 0; The modulation phase of the signal remains unchanged throughout the complete symbol period; Indicates the initial phase; j represents sqrt(-1), where sqrt is the square root; n0(n) follows N(0,σ) 2 Additive Gaussian white noise.

[0015] Step 2, signal visualization construction, is implemented as follows:

[0016] Existing deep learning models are primarily designed to solve image classification problems, requiring the conversion of baseband sequences into constellation diagram images. The baseband sequence can be considered as the in-phase component s of the signal itself. I (n) and orthogonal component s Q (n) constitutes, that is:

[0017]

[0018] Where Re[s(n)] represents the real part information of the baseband sequence, Im[s(n)] represents the imaginary part information of the baseband sequence, Re[] represents taking the real part, and Im[] represents taking the imaginary part.

[0019] First, the baseband sequence is preprocessed by normalization. Then, Python is used to plot the preprocessed baseband sequence as a constellation diagram using the scatter function. Finally, the constellation diagram is converted into a three-channel grayscale image.

[0020] Step 3, image serialization and position encoding, is implemented as follows:

[0021] The input constellation diagram is represented as follows Where H represents the height of the input constellation image, W represents the width of the input constellation image, and C represents the number of channels in the input constellation image. The input constellation image is divided into N P×P patches with non-overlapping pixels. in:

[0022]

[0023] The location information for each patch block is: Each P 2 The C-dimensional patch vectors are then subjected to linear transformations and right-multiplied by matrices. The resulting D-dimensional vector. The global feature of the constellation diagram is obtained by calculating the mean of the D-dimensional matrix obtained by right multiplying each patch, denoted as . The encoder Transformer input sequence obtained after position encoding of each patch block can be represented as:

[0024]

[0025] Step 4, the Transformer encoder, is implemented as follows:

[0026] The encoder of a traditional ViT model consists of stacked Transformer network layers. In contrast, the ViT network of this invention, which uses local feature window fusion, extracts global and local feature information in its l-th layer through a Global Feature Extraction (GFE) module and a Local Feature Extraction (LFE) module, respectively. Here, the output features of the (l-1)th layer are represented as... Let the final output sequence of the encoder be...

[0027] 4-1. Global Feature Extraction (GFE) Module

[0028] The output of the l-th Transformer network layer is fed into another Transformer network layer to obtain the feature vector. Pick g, as a global feature component of the constellation graph class .

[0029] 4-2. Local Feature Extraction (LFE) Module

[0030] Traditional ViT models offer advantages such as parallel training and full utilization of global information, but they neglect the impact of correlations between local information on classification performance. The improved ViT model of this invention adds a Local Feature Extraction (LFE) layer after the (l-1)th Transformer layer, fully utilizing the correlations between local features through an M-layer feature fusion layer. The LFE module first pools the output of the l-th Transformer network layer and then... The combination yields sequence G0, which is:

[0031]

[0032] In the formula, Pool represents the pooling layer, and average pooling is used here; Concate represents... The input sequence G0 is concatenated with the local feature components output from layer l-1 of the network; it then passes through layer M of feature fusion layers. The output sequence of each feature fusion layer... It can be represented as:

[0033]

[0034] Splice represents the local feature window fusion module; Conv represents a 1*1 convolutional layer; P Mask represents the mask matrix; ReLU represents the selected activation function, and m is the m-th feature fusion layer.

[0035] The Splice local feature window fusion module proposed in this invention fully integrates feature information into each local feature block through stitching, allowing each local feature block to carry the content contained in other image blocks, thus fully reflecting the correlation characteristics between local features. The Splice local feature window fusion module is specifically as follows: assuming the number of local features after pooling is K, and each local feature is decomposed into D blocks, during the feature fusion process of the m-th layer, the j1 channel information G of the (i1+j1)%(K+1) block features is sequentially... m-1 ((i1+j1)%(K+1),j1) moves to the j1 channel position G' of the i1th feature block. m-1 (i1, j1), % represents the remainder, and the shift process is as follows:

[0036] G' m-1 (i1,j1)=Gm-1 ((i1+j1)%(K+1),j1)i1=0,...,K; j1=0,...,D-1 (7)

[0037] Among them, G' m-1 (i1,j1) represents the feature sequence after the shift. m-1 Each feature block carries feature information from other blocks, which makes up for the shortcomings of traditional ViT models that ignore the impact of the correlation between local information on classification performance.

[0038] The output sequence G after the M-layer feature fusion layer M The output incorporating local features is obtained through the Transformer in the Local Feature Extraction (LFE) layer. Selecting features Local feature components as encoder output

[0039] Step 5 is as follows:

[0040] 5.1. Classification Result Prediction

[0041] The classifier consists of a multilayer perceptron (MLP) and a softmax function; the global feature components g are obtained from the global feature extraction (GFE) layer. class Local feature components obtained by Local Feature Extraction (LFE) The MLP prediction results will be input separately.

[0042] 5-2. Loss Function:

[0043] As the signal-to-noise ratio decreases, noise significantly impacts baseband information. Converging constellation points gradually diverge due to amplified noise power, leading to a marked increase in the similarity between constellation diagrams of different modulation types that were originally quite distinct. This increases the difficulty of classification. This invention employs cross-entropy loss L in its loss function design. cross-entropy and triplet loss L triplet Joint training method. The loss expression is:

[0044] L = L cross-entropy +L triplet (8)

[0045] For multi-class image classification tasks, cross-entropy loss is often used to learn sample features. Cross-entropy loss L... cross-entropy The mathematical expression is:

[0046]

[0047] In the formula, N pi n represents the number of images in a batch. pif represents the number of signal types. i Let y represent the eigenvector of the i-th constellation diagram. i f i The corresponding real category, W j Let b represent the j-th column vector of matrix W. j This indicates the bias term.

[0048] Since cross-entropy loss only constrains the distance between constellation diagrams of signals with different modulation types and does not consider the impact of the distance between constellation diagrams of signals with the same modulation type on training, triplet loss and cross-entropy loss are added to the network for joint supervision. The network model based on triplet loss can, during training, make the spatial distance between anchors of the same modulation type and positive features as close as possible, and the distance between anchors of different modulation types and negative features as far as possible. This allows for better learning of subtle feature differences between similar constellation diagrams in low signal-to-noise ratio environments. The triplet loss L... triplet The mathematical expression is:

[0049]

[0050] In the formula, batch_size represents the number of images processed in a batch; Represents a random baseline sample in a training batch. Feature mapping; Indicates and Random positive samples in the same batch Feature mapping; Indicates and Random negative samples in the same batch Feature mapping; and α and β represent the Euclidean distance between the positive and negative sample features and the baseline sample features, respectively; α is a positive constant, and the distance between positive and negative samples can be controlled by changing the value of α. 0.2 was used in the experiment; + indicates that when the value in [] is greater than 0, the value is taken as the loss, otherwise the loss is 0.

[0051] 5-3 Signal classification results:

[0052] After cross-entropy loss L cross-entropy and triplet loss L triplet The supervisory network is trained jointly, and the classification results of communication signals with different modulation types are finally output.

[0053] The beneficial effects of this invention are as follows:

[0054] 1. To address the problem that noise interferes with signal distribution characteristics in low signal-to-noise ratio environments, leading to classification difficulties.

[0055] 2. Solve the problem of poor classification performance of high-order signals.

[0056] 3. The signal constellation diagram, after network enhancement, provides clearer extractable features. The addition of a local feature window fusion module fully enhances the correlation between local features, while retaining the advantage of the original ViT network in fully utilizing global information, thus achieving accurate differentiation between different modulation types.

[0057] 4. Under different signal-to-noise ratio environments, compared with AlexNet, ResNet-18, ResNet-50 and ViT networks, the network of this invention significantly improves the classification accuracy.

[0058] In summary, this invention features high recognition rate and strong robustness, and maintains a high recognition rate even at low signal-to-noise ratios. Attached Figure Description

[0059] Figure 1 Three-channel images of eight modulation types of signals at different signal-to-noise ratios;

[0060] Figure 2 The structure of a signal classification model based on local feature window fusion ViT network;

[0061] Figure 3 Transformer network layer structure;

[0062] Figure 4 Feature fusion layer based on local feature window fusion of ViT network;

[0063] Figure 5 Image patch movement process in the local feature window fusion module;

[0064] Figure 6 Visualization of the features of signals constructed into images by different deep learning networks;

[0065] Figure 7 Confusion matrix of classification results for signals with a signal-to-noise ratio of -2dB by different networks;

[0066] Figure 8 The classification results of five network models with different signal-to-noise ratios for 16QAM and 64QAM;

[0067] Figure 9 Classification results of five network models under different signal-to-noise ratios. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. The technical solution adopted by this invention is as follows:

[0069] 1. Eight modulation types, including {BPSK, QPSK, 8PSK, 16APSK, 32APSK, 16QAM, 32QAM, 64QAM}, were obtained through real receivers and simulations, and preprocessing such as carrier synchronization and symbol synchronization was performed on the signals.

[0070] 2. Representing the signal as a constellation diagram can lead to information loss during data conversion. To minimize conversion loss and more intuitively reflect the correlation between signal symbols and the integrity of signal characteristics, this invention uses Python to generate a three-channel grayscale image with a resolution of 224×224. Figure 1 The images show three-channel representations of eight modulation types at three different signal-to-noise ratios (SNRs): 0dB, 10dB, and 20dB. Observation reveals that as the SNR decreases, the signal constellation diagram becomes blurred, and noise affects the classification results.

[0071] 3. The constellation diagram of the normalized digital baseband signal of the input image is represented as follows: Where H represents the height of the input image, W represents the width of the input image, and C represents the number of channels in the input image. The input image is divided into N P×P non-overlapping patch blocks.

[0072] Figure 2 The boxes numbered 1 to N indicate the location information of each patch block. The right box indicates each P 2 The C-dimensional patch vectors are then subjected to linear transformations and right-multiplied by matrices. The resulting D-dimensional vector, with boxes marked with * representing global features of the entire constellation diagram.

[0073] 4. The Transformer encoder first passes through l Transformer network layers, then the output features of the l-th Transformer network layer are input into two branches. One branch passes through the Global Feature Extraction (GFE) module and one Transformer network layer; the other branch passes through the Local Feature Extraction (LFE) module and one Transformer network layer. These two branches output the global feature components g. class With local feature components Figure 3The specific structure of each Transformer network layer consists of multiple layer normalizations and a multi-head self-attention mechanism.

[0074] Global Feature Extraction (GFE) module

[0075] The Global Feature Extraction (GFE) module takes the output of the l-th Transformer network layer and feeds it into another Transformer network layer to obtain the feature vector. Pick g, as a global feature component of the constellation graph class .

[0076] Local Feature Extraction (LFE) module

[0077] The improved ViT model of this invention adds a Local Feature Extraction (LFE) layer after the (l-1)th Transformer layer, and fully utilizes the correlation between local features through an M-layer feature fusion layer. The local feature portions in the output sequence of the l-th Transformer network layer are pooled, and the pooled vector is then... The input sequence G0 of the feature fusion layer is obtained by combining the sequences.

[0078] The input sequence G0 will pass through an M-layer feature fusion layer, as shown in the example. Figure 4 As shown. Figure 4 In the diagram, Splice represents the local feature window fusion module; Conv represents a 1*1 convolutional layer; P Mask Represents the mask matrix; ReLU represents the selected activation function. The output sequence after the M-layer feature fusion layer is:

[0079] The local feature fusion module integrates feature information into each local feature block through a specific pattern of stitching. The specific moving process is as follows: Figure 5 As shown, each feature block after being moved carries feature information from the other blocks.

[0080] The output sequence G after the M-layer feature fusion layer M The output incorporating local features is obtained through the Transformer in the Local Feature Extraction (LFE) layer. Selecting features Local feature components as encoder output

[0081] 5. Global feature components g obtained from the global feature extraction layer class Local feature components obtained from local feature extraction The MLP prediction results will be input separately.

[0082] For multi-class image classification tasks, cross-entropy loss L is often used. cross-entropy The network learns from the sample features, but since the cross-entropy loss only constrains the distance between constellation diagrams of signals with different modulation types and does not consider the impact of the distance between constellation diagrams of signals with the same modulation type on the training, a triplet loss L is added to the network. triplet With cross-entropy loss L cross-entropy Joint supervision will be conducted.

[0083] Network models based on triplet loss can make the spatial distance between the features of the reference samples and positive samples of the same modulation type as close as possible during training, and the distance between the features of the reference samples and negative samples of different modulation types as far as possible. This allows them to better learn the subtle feature differences of similar constellation diagrams in low signal-to-noise ratio environments.

[0084] Example:

[0085] 1. Dataset

[0086] The experiment used eight signal modulation types: BPSK, QPSK, 8PSK, 16APSK, 32APSK, 16QAM, 32QAM, and 64QAM. Signal samples were obtained using a real receiver and through simulation, and preprocessing was performed, including carrier synchronization and symbol synchronization. The baseband signal dataset had a signal-to-noise ratio range of -4 to 6 dB with a step size of 2. The complex baseband signal samples contained 1000 symbols. The complex signal samples were represented as normalized digital baseband signal constellation diagrams, with images being 224×224 resolution three-channel grayscale images. For each modulation type, 4000 images were used as the training set and 1000 images as the test set, with corresponding type labels assigned to the image samples. The training set was enhanced using affine transformations such as image flipping and random angle rotation to ensure the network's robustness to the constellation diagram's feature scale and orientation.

[0087] 2. Experimental Environment

[0088] The DL model training process uses GPU to train the model to accelerate the data computation rate. The GPU used is NVIDIA TITAN RTX 3090, CUDA version is 11.4, the operating system is 64-bit Ubuntu 16.04, the deep learning framework used is PyTorch 1.8.0, and the deep learning training model is based on Python 3.7.9.

[0089] 3. Network Configuration

[0090] The deep learning network inputs a Transformer encoder with N = 196 patches, l = 12 Transformer layers, K = 4 local features after pooling, and M = 2 feature fusion layers. An Adaptive Moment Estimation (ADAM) optimizer is used for iterative optimization. After multiple experiments, the initial learning rate is set to 0.0003, the decay coefficient to 0.0005, the training epochs to 100, and the batch size to 64.

[0091] 4. Performance Simulation

[0092] In addition to the network proposed in this invention, this example also uses the AlexNet network from "Modulation Classification Based on Signal Constellation Diagrams and Deep Learning" (Reference 1), the ResNet-18 network from "Modulation classification of MQAM signals based on gradient colorconstellation and deep learning" (Reference 2), the ResNet-50 network from "Deep residual learning for image recognition" (Reference 3), and the ViT network from "An image is worth 16x16 words: Transformers for image recognition at scaleViT" (Reference 4).

[0093] 4-1. Feature Visualization

[0094] Under a -2dB signal-to-noise ratio environment, the feature visualization heatmap of an image constructed from five high-order amplitude-phase modulation signals in the signal set, using the ViT network and ResNet-18 network fusion method proposed in this invention, is shown below. Figure 6 As shown, different colors represent the degree of attention paid to regional features. Darker colors indicate the regions where signal features are of particular interest during classification. The five rows from top to bottom are BPSK, 8PSK, 32APSK, 32QAM, and 64QAM. The four columns from left to right are the original image, ResNet-18 network, traditional ViT network, and ViT network with local feature window fusion.

[0095] Observing the original images reveals that in low signal-to-noise ratio environments, the distribution characteristics of constellation images are not obvious, and several types of constellation images are quite similar. Figure 6 It can be observed that ResNet-18 and ViT networks suffer from issues such as a small region of interest and background interference. For feature visualization images of the same image, the ViT network has a smaller background interference region compared to ResNet-18, but interference still exists in classification. The network proposed in this invention fully combines the characteristics of global and local features, pays close attention to the distribution characteristics of the constellation diagram, effectively suppresses background influence, and achieves better classification results than the other two networks.

[0096] 4-2. Confusion Matrix

[0097] A signal set containing eight signal types under a -2dB signal-to-noise ratio environment was selected as the research object. The classification confusion matrix of the model was statistically analyzed under three network training conditions. The test set samples for each modulation type were 1000. Figure 7 From left to right, the images show the confusion matrices of the classification results for ResNet-18, ViT network, and ViT network with local feature window fusion. Figure 7 The three sets of confusion matrices reveal that the ViT network, which integrates local feature windows according to this invention, has a better classification effect for signals with more complex amplitude and phase information compared to the ResNet-18 network and the ViT network.

[0098] Figure 8 The classification results of two high-order amplitude-phase hybrid modulation 16QAM and 64QAM under different signal-to-noise ratio (SNR) environments were compared using five different models. The left figure shows the classification results of 16QAM under different SNR, and the right figure shows the classification results of 64QAM under different SNR. It can be observed that in low SNR environments, the ViT network model with local feature window fusion proposed in this invention has better performance than the other four network models for classifying complex modulation signals with rich amplitude and phase information.

[0099] 4-3. Classification Performance Analysis

[0100] Assuming the training and test sets are identical, the average classification performance of eight amplitude-phase modulation signals under different signal-to-noise ratios (SNRs) is statistically analyzed. The classification results of the five network models under different SNRs are as follows: Figure 9 As shown.

[0101] Figure 9As can be seen, for the classification of eight signal classes under a signal-to-noise ratio (SNR) of -4dB, the classification accuracy using AlexNet is 67.25%, using ResNet-18 is 68.14%, using ResNet-50 is 69.73%, using ViT is 70.97%, and using the ViT network with local feature window fusion of this invention achieves 72.19%. This represents an improvement of 4.94%, 4.05%, 2.46%, and 1.22% compared to AlexNet, ResNet-18, ResNet-50, and ViT, respectively. Data analysis shows that under different SNR environments, the ViT network with local feature window fusion of this invention significantly improves classification accuracy compared to AlexNet, ResNet-18, ResNet-50, and ViT.

[0102] Finally, it should be noted that the purpose of disclosing the embodiments is to help further understand the present invention. However, those skilled in the art will understand that various substitutions and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the embodiments, and the scope of protection of the present invention is defined by the claims.

Claims

1. A communication signal classification method based on a local feature window fusion ViT network, characterized in that, The method comprises the following steps: Step 1: Obtain communication signal samples under different signal-to-noise ratios through a real receiver and simulation simulation, perform carrier synchronization and symbol synchronization processing, and then convert into a baseband sequence with additive white Gaussian noise; Step 2: Preprocess the baseband sequence through normalization, and convert the preprocessed baseband sequence into a constellation diagram; Step 3: Divide the constellation diagram into N patches, encode the position of each patch, and obtain an input sequence of a transform encoder; The specific process of the position encoding is as follows: The constellation map is represented as where H represents the height of the input constellation map, W represents the width of the input constellation map, C represents the number of channels of the input constellation map, and the input image is cut into N P x P size patches without pixel overlap: i = 1, 2,..., N, where: The position information of each patch is Each P 2 The patch vectors of C dimensions are respectively subjected to linear transformation right multiplication matrix The D-dimensional vector is obtained; the global feature of the constellation diagram is obtained by stacking and calculating the mean value of the D-dimensional matrix obtained by right multiplication of each patch, denoted as The position encoding of each patch block is obtained: z0 is the input sequence of the encoder transform; Step 4, input the input sequence into the Transformer encoder, first pass through the l-th layer of the Transformer network layer, and then input the output features of the l-th layer of the Transformer network layer into two branches: one branch passes through the global feature extraction GFE module and a layer of the Transformer network layer; the other branch passes through the local feature extraction LFE module and a layer of the Transformer network layer, and the two branches respectively output the global feature component g class and the local feature component In step 4, the output feature representation of the l-th layer of the Transformer network is denoted as: The global feature extraction GFE module inputs the output of the l-th layer of the Transformer network layer into another Transformer network layer again to obtain a feature vector Take As a constellation global feature component g class ; The local feature extraction LFE module first pools the output of the lth layer of the Transformer network, and then combines it with G0, that is: In the formula, Pool represents a pooling layer; and Concate represents a splicing operation. The G0 will pass through M layer feature fusion layers, and the output sequence of each feature fusion layer is is represented as: G m = ReLU(Conv(Splice(G m-1 )) ⊙ P Mask + G m-1 ), m = 1, 2,..., M (5) Splice denotes a local feature window fusion module; Conv denotes a convolutional layer; P Mask denotes a mask matrix; ReLU denotes an activation function; m is the mth feature fusion layer; Output sequence G through the M-layer feature fusion layer M Input a layer of Transformer network layer to obtain an output fused with local features Select features As a local feature component; Step 5, the global feature component g class with the local feature component Input the classifier and construct the loss function to train the Transformer encoder to obtain the communication signal classification result.

2. The communication signal classification method based on the local feature window fusion ViT network according to claim 1, characterized in that, In step 1: the set of communication signal samples includes eight modulation types of binary phase shift keying (BPSK), quadrature phase shift keying (QPSK), 8-phase shift keying (8PSK), 16-amplitude phase shift keying (16APSK), 32-amplitude phase shift keying (32APSK), 16-quadrature amplitude modulation (16QAM), 32-quadrature amplitude modulation (32QAM) and 64-quadrature amplitude modulation (64QAM), and the general model s(n) of the eight modulation types of communication signals is: where a(n) represents the amplitude of the nth symbol; f c is the center frequency of the signal, and if the signal has been carrier-synchronized, f c = 0. is the modulation phase of the signal, which remains constant over a complete symbol period; represents the initial phase; j represents the square root of -1; n0(n) is additive white Gaussian noise with N(0, σ 2 ).

3. The communication signal classification method based on the local feature window fusion ViT network according to claim 2, characterized in that, In step 2, the baseband sequence is composed of the in-phase component s I (n) and the quadrature component s Q (n) of the communication signal, i.e.: Wherein, Re[s(n)] represents the real part information of the baseband sequence, Im[s(n)] represents the imaginary part information of the baseband sequence, Re[] represents taking the real part, and Im[] represents taking the imaginary part.

4. The communication signal classification method based on the local feature window fusion ViT network according to claim 3, characterized in that, In step 2, the specific operation of converting the preprocessed baseband sequence into a constellation diagram is as follows: using Python, the preprocessed baseband sequence is converted into a constellation diagram form by using a scatter function, and then the constellation diagram is converted into a three-channel grayscale image.

5. The communication signal classification method based on the local feature window fusion ViT network according to claim 4, characterized in that, The specific operation of the local feature window fusion module Splice is as follows: assuming that the number of pooled local features is K, each local feature is decomposed into D blocks, and the j1 channel information G m-1 ((i1+j1)%(K+1),j1) of the (i1+j1) % (K+1) block features is moved to the j1 channel position G m-1 (i1,j1) of the i1 block feature, % represents the remainder, and the moving process is: G' m-1 (i1,j1) = G m-1 ((i1+j1)%(K+1),j1)i1=0,…,K; j1=0,…,D-1 (5) where G' m-1 (i1,j1) denotes the shifted feature sequence.

6. The communication signal classification method based on the local feature window fusion ViT network according to claim 5, characterized in that, In step 5, the classifier is composed of a multilayer perceptron (MLP) and a softmax function; The loss function expression is: L = L cross-entropy + L triplet (5) where L cross-entropy is the cross-entropy loss, L triplet is the triplet loss.

Images