Communication signal modulation identification method based on improved ViT model and computer device

Abstract

The invention relates to the field of communication signal modulation identification, in particular to a communication signal modulation identification method based on an improved ViT model and a computer device. The method comprises the following steps: collecting communication signal data of different modulation types, sampling rates and bandwidths, and preprocessing the data to generate a modulation identification signal sample set; performing normalization processing on the modulation identification signal sample set, and dividing the modulation identification signal sample set into a training sample set, a verification sample set and a test sample set according to different data fields; constructing an improved ViT model, and performing feature extraction on the training sample set through the improved ViT model; features extracted by the improved ViT model are input into a classifier, back propagation is used for the improved ViT model and the classifier to iteratively optimize parameters of the improved ViT model till a loss function does not decrease any more, and finally modulation recognition is conducted on communication signals through the improved ViT model. The method is suitable for communication signal modulation identification.

Description

Technical Field

[0001] The present invention relates to the field of communication signal modulation recognition, and in particular to a communication signal modulation recognition method and a computer device based on an improved ViT model. Background Art

[0002] As wireless communication environments become increasingly complex, with diverse signal types, increasing interference levels, and rapidly changing conditions, the demand for advanced modulation recognition techniques that can operate effectively in this harsh and dynamic environment is growing. These techniques must be robust enough to handle low signal-to-noise ratios, multipath fading, and other challenging conditions while maintaining high accuracy and reliability. Therefore, in order to ensure the efficient and secure operation of communication systems, it is becoming increasingly important to develop modulation recognition methods that are not only accurate but also able to adapt to modern wireless environments.

[0003] Current modulation recognition technologies are mainly divided into two categories: likelihood-based and feature-based. Likelihood-based automatic modulation classification methods treat signal parameters as variables and convert the problem into hypothesis testing. Although these methods are relatively comprehensive in theory, they are very sensitive to signal parameters and involve complex calculations with a large amount of computation.

[0004] On the other hand, feature-based methods focus on constructing and extracting meaningful features from signals, which are then used for classification. Unlike likelihood-based methods, feature-based methods are more versatile and can identify a wider range of signal types even with limited prior knowledge. However, these methods rely heavily on the effectiveness of the feature extraction process; if the features are not well constructed, the accuracy of signal classification will be greatly reduced. At the same time, when faced with different modulation modes and channel conditions, specific feature extraction techniques need to be developed for each situation, which not only consumes a lot of manpower and time, but also significantly reduces the recognition accuracy under the influence of noise and interference, resulting in insufficient robustness of the system. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a communication signal modulation recognition method and a computer device based on an improved ViT (Vision Transformer, image classification based on Transformer architecture) model, thereby improving the communication signal modulation recognition accuracy.

[0006] The present invention adopts the following technical solutions to achieve the above-mentioned purpose. In a first aspect, the present invention provides a communication signal modulation identification method based on an improved ViT model, comprising: S1. Collect communication signal data of different modulation types, sampling rates, and bandwidths, and generate a modulation identification signal sample set after preprocessing the data; S2, normalizing the modulation identification signal sample set, and dividing the modulation identification signal sample set into a training sample set, a verification sample set, and a test sample set according to different data domains; S3, constructing an improved ViT model, and extracting features from the training sample set through the improved ViT model; S4. Input the features extracted by the improved ViT model into the classifier for modulation recognition, and use back propagation to iteratively optimize the parameters of the improved ViT model until the loss function no longer decreases. Finally, the communication signal is modulated through the improved ViT model.

[0007] Furthermore, step S1 specifically includes: The modulated signal data of various modulation types, sampling rates and bandwidths are collected through the receiver. After the noise is removed by a low-pass filter, the received modulated signal data is divided into signal segments of equal length, and each signal segment is taken as a sample, as a modulated signal sample. Multiple modulated signal samples constitute a modulation identification signal sample set.

[0008] Furthermore, step S2 specifically includes: The modulation identification signal sample set is normalized using the zero-mean normalization formula. 80% of the samples are randomly selected from different data domains to form a training sample set, 10% of the samples form a verification sample set, and the remaining 10% of the samples form a test sample set.

[0009] Furthermore, in step S3, the specific process of constructing the improved ViT model includes: First, the input modulated signal sample data shape is transformed into (2, L, 1), where 2 represents the two different channels of IQ data, L represents the input modulated signal length, and 1 is the expanded dimension, the purpose is to make the input shape adapt to the improved ViT model. Then, the input modulated signal is divided into several patches in the length dimension, and concatenated in the channel dimension, and the last two dimensions are flattened to obtain ,in, Represents the number of patches. Represents the length of each patch, Indicates divisibility; get After that, a cls_token is spliced ​​in front of it for classification, and its shape is , the concatenated input , then, positionally encode the input to construct a learnable position vector and compare it with the input Add together to get the final input , as the input of the improved ViT; Then, the encoder module is constructed. The specific structure includes: multi-head attention module, layer normalization, random drop layer and fully connected layer. First, we use the multi-head attention module, which is composed of several self-attention modules. Each self-attention module contains , and Three matrices; First, Multiplying these three matrices respectively, we get , , ,in, Represents Query (query matrix), represents Key (key matrix), represents Value (value matrix). Then, Pooling is performed to obtain the proxy matrix ,in, represents the dimension of the proxy matrix, and secondly, and Multiply and perform Softmax operation to obtain the proxy matrix Weighted query matrix ,in, represents the matrix transpose, and Multiply and perform Softmax operation to get the proxy matrix Weighted bond matrix ,Will and Multiply them together to get the weighted key matrix The value matrix after matching Finally, and Multiply and pass through a random dropout layer to get the attention feature ; Then, the attention feature Normalized by layers and compared with the input Add and get the output , then Input the fully connected layer and randomly discard it to get the output Finally, and Add together to get the output of the multi-head attention module , stack several encoder modules and extract features from the input modulation signal to obtain the final output features .

[0010] Furthermore, step S4 specifically includes: For the final output features First, take the cls_token and flatten it to get , and input it into the fully connected layer 1 to obtain the output features of the fully connected layer 1 ,in, represents the dimension of the fully connected layer 1, and then Input the batch normalization layer and use the ReLU function for activation to obtain the activated features , then, Input the fully connected layer 2 to get the output features of the fully connected layer 2 ,in, represents the number of modulation categories. Finally, Through Softmax, the probability that the signal sample belongs to different modulation modes is obtained; For the obtained , here the cross entropy loss is used for network optimization as follows: ,in, is the number of modulation categories, is the true label, The labels predicted by the network; The network parameters are optimized until the set number of optimizations is reached. After the optimization is completed, the communication signal is modulated and identified using the improved ViT model.

[0011] In a second aspect, the present invention provides a computer device, including a memory, wherein the memory stores program instructions, and when the program instructions are executed, the communication signal modulation identification method based on the improved ViT model described above is executed.

[0012] The beneficial effects of the present invention are: The present invention pre-processes the collected communication signals of various modulation types, and inputs them into the pre-established improved ViT model, iterates the rounds, completes the overall training, and finally outputs the modulation mode of the signal. The improved ViT model includes the Q, K, V matrices in the original ViT model and the encoding layer composed of the A matrix introduced separately, followed by the cascaded fully connected layer, ReLU layer, batch normalization layer and output layer. The present invention reduces the computational complexity of the overall network, improves the inference speed of the model, and ultimately improves the accuracy of modulation recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a flow chart of a communication signal modulation identification method based on an improved ViT model provided by an embodiment of the present invention; Figure 2is a schematic diagram of the structure of the self-attention module of the improved ViT model provided by an embodiment of the present invention; Figure 3 It is a process flow chart of a classifier provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0014] To make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0015] The present invention provides a communication signal modulation recognition method based on an improved ViT model, such as Figure 1 As shown, specifically including: S1. Collect communication signal data of different modulation types, sampling rates, and bandwidths, and generate a modulation identification signal sample set after preprocessing the data; Specifically, the present invention collects multiple modulation type signals with different modulation types, different bandwidths and different sampling rates through a receiver, including but not limited to digital modulation types 4ASK, 8ASK, BPSK, QPSK, OQPSK, PI4QPSK, 8QAM, 16QAM, GMSK and analog modulation types FM, AM-SSB-WC, AM-SSB-SC, AM-DSB-WC, AM-DSB-SC.

[0016] Then, the noise is removed by a low-pass filter, and the received modulated signal data is cut into signal segments of equal length. Each signal segment is taken as a sample, that is, a modulated signal sample. Multiple modulated signal samples constitute a modulation identification signal sample set.

[0017] S2, normalizing the modulation identification signal sample set, and dividing the modulation identification signal sample set into a training sample set, a verification sample set, and a test sample set according to different data domains; Specifically, the modulation recognition signal sample set is normalized using the zero-mean normalization formula, and 80% of the samples are randomly selected from different data domains to form a training sample set, 10% of the samples form a verification sample set, and the remaining 10% of the samples form a test sample set.

[0018] The zero mean normalization formula is as follows: ; Indicates Signal samples The sample values ​​after normalization are is the mean of a single sample data, is the standard deviation of a single sample data.

[0019] The calculation formula is as follows: ; Represents the total number of sampling points for a single sample of data.

[0020] S3, constructing an improved ViT model, and extracting features from the training sample set through the improved ViT model; Based on the training sample set, the corresponding feature vector is obtained through the improved ViT model; The specific process of constructing the improved ViT model includes: First, the input modulated signal sample data shape is transformed into (2, L, 1), where 2 represents the two different channels of IQ data, L represents the input modulated signal length, and 1 is the expanded dimension, the purpose is to make the input shape adapt to the improved ViT model. Then, the input modulated signal is divided into several patches in the length dimension, and concatenated in the channel dimension, and the last two dimensions are flattened to obtain ,in, Represents the number of patches. Represents the length of each patch, Indicates divisibility; get After that, a cls_token is spliced ​​in front of it for classification, and its shape is , the concatenated input , then, positionally encode the input to construct a learnable position vector and compare it with the input Add together to get the final input , as the input of the improved ViT; Then, the encoder module is constructed. The specific structure includes: multi-head attention module, layer normalization, random drop layer and fully connected layer. First, through the multi-head attention module, the multi-head attention module is composed of several self-attention modules. Figure 2 As shown, each self-attention module contains , and Three matrices; First, Multiplying these three matrices respectively, we get , , ,in, stands for Query, Represents Key, represents Value. Then, Pooling is performed to obtain the proxy matrix ,in, represents the dimension of the proxy matrix, and secondly, and Multiply and perform Softmax operation to get the proxy matrix Weighted query matrix ,in, represents the matrix transpose, and Multiply and perform Softmax operation to get the proxy matrix Weighted bond matrix ,Will and Multiply them together to get the weighted key matrix The value matrix after matching Finally, and Multiply and pass through a random dropout layer to get the attention feature ; Then, the attention feature Normalized by layers and compared with the input Add and get the output , then Input the fully connected layer and randomly discard it to get the output Finally, and Add together to get the output of the multi-head attention module , stack several encoder modules and extract features from the input modulation signal to obtain the final output features .

[0021] S4. Input the features extracted by the improved ViT model into the classifier, and use back propagation to iteratively optimize the parameters of the improved ViT model until the loss function no longer decreases. Finally, the communication signal is modulated and recognized through the improved ViT model.

[0022] For the final output features ,First, cls_token is taken out and flattened, and input into the classifier for ,modulation recognition, and the model parameters are updated using ,backpropagation for multiple iterations until the loss function no longer decreases.

[0023] like Figure 3 As shown, the classifier processing flow is as follows: For the final output features First, take the cls_token and flatten it to get , and input it into the fully connected layer 1 to obtain the output features of the fully connected layer 1 ,in, represents the dimension of the fully connected layer 1, and then Input the batch normalization layer and use the ReLU function for activation to obtain the activated features , then, Input the fully connected layer 2 to get the output features of the fully connected layer 2 ,in, represents the number of modulation categories. Finally, Through the Softmax function, the probability that the signal sample belongs to different modulation modes is obtained; For the obtained , here the cross entropy loss is used for network optimization as follows: ,in, is the number of modulation categories, is the true label, The labels predicted by the network; Set the learning rate drop mechanism and early stop mechanism. After the iterative optimization is completed, the trained modulation recognition network model is obtained. The specific setting method is as follows: When the Loss value of the validation set does not decrease within 6 cycles, multiply the learning rate by one third; when the number of training times set by the deep neural network is reached or the Loss value of the validation set does not decrease within 20 cycles, the training of the neural network is completed. The final converged Loss value should be as small as possible. After completing the training of the neural network, input test samples of different modulation types into the trained model to obtain the modulation recognition results; and compare the recognition results with the true categories of the test samples to statistically calculate the recognition accuracy.

[0024] The above is only a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be modified within the scope of the concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art shall not deviate from the spirit and scope of the present invention, and shall be within the scope of protection of the claims attached to the present invention.

Claims

1. A communication signal modulation identification method based on an improved ViT model, characterized in that: include: S1. Collect communication signal data of different modulation types, sampling rates, and bandwidths, and generate a modulation identification signal sample set after preprocessing the data; S2, normalizing the modulation identification signal sample set, and dividing the modulation identification signal sample set into a training sample set, a verification sample set, and a test sample set according to different data domains; S3, constructing an improved ViT model, and extracting features from the training sample set through the improved ViT model; S4. Input the features extracted by the improved ViT model into the classifier, and use back propagation to iteratively optimize the parameters of the improved ViT model until the loss function no longer decreases. Finally, the communication signal is modulated and recognized through the improved ViT model.

2. The communication signal modulation identification method based on the improved ViT model according to claim 1 is characterized in that: Step S1 specifically includes: The modulated signal data of various modulation types, sampling rates and bandwidths are collected through the receiver. After the noise is removed by a low-pass filter, the received modulated signal data is divided into signal segments of equal length, and each signal segment is taken as a sample, as a modulated signal sample. Multiple modulated signal samples constitute a modulation identification signal sample set.

3. The communication signal modulation identification method based on the improved ViT model according to claim 1 is characterized in that: Step S2 specifically includes: The modulation identification signal sample set is normalized using the zero-mean normalization formula. 80% of the samples are randomly selected from different data domains to form a training sample set, 10% of the samples form a verification sample set, and the remaining 10% of the samples form a test sample set.

4. The communication signal modulation identification method based on the improved ViT model according to claim 1 is characterized in that: In step S3, the specific process of constructing the improved ViT model includes: First, the input modulated signal sample data shape is converted to (2, L, 1), where 2 represents the two different channels of IQ data, L represents the input modulated signal length, and 1 is the expanded dimension. Then the input modulated signal is divided into several patches in the length dimension, concatenated in the channel dimension, and the last two dimensions are flattened to obtain ,in, Represents the number of patches. Represents the length of each patch, Indicates divisibility; get After that, a cls_token is spliced ​​in front of it for classification, and its shape is , the concatenated input , then, positionally encode the input to construct a learnable position vector and compare it with the input Add together to get the final input , as the input of the improved ViT; Then, the encoder module is constructed. The specific structure includes: multi-head attention module, layer normalization, random drop layer and fully connected layer. First, we use the multi-head attention module, which is composed of several self-attention modules. Each self-attention module contains , and Three matrices; First, Multiplying these three matrices respectively, we get , , ,in, represents the query matrix, represents the bond matrix, represents the value matrix, then, Pooling is performed to obtain the proxy matrix ,in, represents the dimension of the proxy matrix, and secondly, and Multiply and perform Softmax operation to get the proxy matrix Weighted query matrix ,in, represents the matrix transpose, and Multiply and perform Softmax operation to get the proxy matrix Weighted bond matrix ,Will and Multiply them together to get the weighted bond matrix The value matrix after matching Finally, and Multiply and pass through a random dropout layer to get the attention feature ; Then, the attention feature Normalized by the layer and compared with the input Add and get the output , then Input the fully connected layer and randomly discard it to get the output Finally, and Add together to get the output of the multi-head attention module , stack several encoder modules and extract features from the input modulation signal to obtain the final output features .

5. The communication signal modulation identification method based on the improved ViT model according to claim 4 is characterized in that: Step S4 specifically includes: For the final output features First, take the cls_token and flatten it to get , and input it into the fully connected layer 1 to obtain the output features of the fully connected layer 1 ,in, represents the dimension of the fully connected layer 1, and then Input the batch normalization layer and use the ReLU function for activation to obtain the activated features , then, Input the fully connected layer 2 to get the output features of the fully connected layer 2 ,in, represents the number of modulation categories. Finally, Through Softmax, the probability that the signal sample belongs to different modulation modes is obtained; For the obtained , here the cross entropy loss is used for network optimization as follows: ,in, represents the cross entropy loss function, is the number of modulation categories, is the true label, The labels predicted by the network; The network parameters are optimized until the set number of optimizations is reached. After the optimization is completed, the communication signal is modulated and identified using the improved ViT model.

6. A computer device comprising a memory, wherein the memory stores program instructions, characterized in that: When the program instructions are executed, the communication signal modulation identification method based on the improved ViT model as described in any one of claims 1 to 5 is executed.

Images