Signal Modulation Recognition Method Based on Recursive Convolutional Network and Attention Mechanism

Abstract

The present invention provides a signal modulation recognition method based on a recursive convolutional network and an attention mechanism, comprising the following steps: sampling each communication signal to generate a data set; performing data augmentation on the data set; establishing a signal modulation recognition model, the signal modulation recognition model including a recursively arranged convolutional network and a multi-scale feature fusion module, the recursively arranged convolutional network and the multi-scale feature fusion module being used to extract feature information of communication signal data; training the signal modulation recognition model using the data set after data augmentation, and further optimizing the signal modulation recognition model using a residual neural network to obtain the optimized signal modulation recognition model; building a classifier for the optimized signal modulation recognition model. This method can fully extract the feature information of the signal and can improve the calculation speed and recognition effect of traditional signal modulation recognition methods.

Description

Technical Field

[0001] The present invention relates to the technical field of signal modulation recognition, and particularly to a signal modulation recognition method based on a recursive convolutional network and an attention mechanism. Background Art

[0002] With the development of communication technologies, the communication environment has become increasingly complex. To meet the needs of a large number of users, there are more and more signal modulation methods. Signal modulation recognition refers to identifying the modulated signal from noise and ensuring demodulation and feature extraction, which plays an important role in fields such as communication. For example, in the civilian field, it can be used to detect illegal radio and monitor whether the parameter configuration of legal radio meets the standards. In addition, signal modulation recognition also plays a crucial role in the cognitive ratio.

[0003] Signal modulation recognition mainly consists of three processes: data preprocessing, feature extraction, and classification decision-making. Among them, data preprocessing estimates the carrier frequency and symbol rate after down-converting the signal, providing several suitable data for subsequent operations; feature extraction is the most important of the three processes, directly affecting the recognition effect, which refers to transforming the original data to extract some features that are easy to classify. However, in traditional signal modulation recognition methods, the extraction of signal feature information is insufficient, the calculation speed is slow, and the recognition effect is poor.

[0004] Therefore, it is necessary to provide a new signal modulation recognition method that can fully extract the feature information of the signal and improve the calculation speed and recognition effect of traditional signal modulation recognition methods. Summary of the Invention

[0005] Technical Problem to be Solved

[0006] Aiming at the above-mentioned drawbacks of the prior art, the present invention provides a signal modulation recognition method based on a recursive convolutional network and an attention mechanism, which can fully extract the feature information of the signal and improve the calculation speed and recognition effect of traditional signal modulation recognition methods.

[0007] Technical Solution

[0008] To achieve the above object, the present invention is realized through the following technical solutions:

[0009] The present invention provides a signal modulation recognition method based on a recursive convolutional network and an attention mechanism, including the following steps:

[0010] S1. Sampling each communication signal to generate a data set;

[0011] S2. Performing data augmentation on the data set;

[0012] S3. Establish a signal modulation recognition model, where the signal modulation recognition model includes a recurrent convolutional network and a multi-scale feature fusion module arranged in sequence. The recurrent convolutional network and the multi-scale feature fusion module are used to extract the feature information of communication signal data;

[0013] S4. Use the dataset after data augmentation to train the signal modulation recognition model, and further optimize the signal modulation recognition model using a residual neural network to obtain the optimized signal modulation recognition model;

[0014] S5. Build a classifier for the optimized signal modulation recognition model.

[0015] Further, the recurrent convolutional network performs layer-by-layer convolution operations on multiple feature tensors in a recursive manner and fuses with a branch having a channel attention mechanism to integrate the multi-scale feature information of the data.

[0016] Further, the multi-scale feature fusion module uses convolution operations with different convolutional kernel sizes and generates a weight vector to act on the input of the signal data.

[0017] Further, the dataset includes a training set and a test set, and the data volume ratio of the training set to the test set is 7:3.

[0018] Further, step S1 further includes:

[0019] Perform data standardization on the data in the dataset:

[0020]

[0021] where x is the original data, μ is the average value of the dataset samples, σ is the standard deviation of the dataset samples, and x * is the data after standardization processing;

[0022] Perform sample alignment on the dataset: Translate the data in the dataset so that its data center of gravity g1 moves to nearby, where the calculation method of the data center of gravity g1 is as follows:

[0023]

[0024] Further, step S2 specifically includes: Using supervised single-sample data augmentation to process the data in the dataset, and performing transposition and merging operations on some signal data feature vectors to expand the dataset.

[0025] Further, the data processing process of the recurrent convolutional network is specifically as follows:

[0026] Perform pooling layer operation on signal data: Divide the input signal data feature tensor into several non - overlapping 2×2 - sized square regions, and calculate the average value of the 4 values in each region;

[0027] Perform the pooling layer operation four times to obtain four identical feature tensors c i , then perform a convolutional operation with a recursive structure, and at the same time introduce another branch. This branch processes the input using a channel attention mechanism, and uses the ReLU activation function and a batch normalization processing layer for output.

[0028] Furthermore, the data processing process of the multi - scale feature fusion module is specifically as follows:

[0029] First, perform batch normalization on the input, then perform convolutional operations on each channel using three convolutional kernels with sizes of 3, 5, and 7 respectively. Use zero padding to ensure that the output two - dimensional sizes are the same, then add them and perform a 1×1 convolutional operation to obtain a weight vector. Finally, multiply it with the input, and apply the weight to the feature tensor to obtain the output.

[0030] Furthermore, the residual neural network includes a convolutional input layer and a residual convolutional layer, and outputs through a fully - connected layer.

[0031] Furthermore, the classifier includes a fully - connected layer and a Softmax layer. Among them, the fully - connected layer uses matrix - vector multiplication to perform feature space transformation and integrate the local feature information extracted by the previous network layers, and multiplies the weight matrix with the input vector and adds a bias, and then classifies through the Softmax layer.

[0032] Beneficial effects

[0033] The present invention provides a signal modulation recognition method based on a recursive convolutional network and an attention mechanism. This method can fully extract the feature information of signals, and can improve the calculation speed and recognition effect of traditional signal modulation recognition methods; furthermore, this method progresses layer by layer in a recursive manner when processing signals, and can effectively mine deep - level feature information; still further, this method uses a channel attention mechanism to fuse channel information, which is helpful for feature extraction; finally, this method performs additional pre - processing on data in the pre - training stage to increase the diversity of communication signal samples. The pre - processed model will no longer be limited by a large number of training samples, and has practical significance in practical applications. Description of the drawings

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0035] Figure 1 Schematic diagram of the steps of a signal modulation recognition method based on a recursive convolutional network and an attention mechanism provided by an embodiment of the present invention;

[0036] Figure 2 Schematic diagram of the model framework of a signal modulation recognition method based on a recursive convolutional network and an attention mechanism provided by an embodiment of the present invention. Detailed implementation manners

[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0038] Refer to Figure 1 and Figure 2 An embodiment of the present invention provides a signal modulation recognition method based on a recursive convolutional network and an attention mechanism, including the following steps:

[0039] S1. Sample each communication signal to generate a data set;

[0040] S2. Perform data augmentation on the data set;

[0041] S3. Establish a signal modulation recognition model, which includes a recursive convolutional network and a multi-scale feature fusion module arranged in sequence. The recursive convolutional network and the multi-scale feature fusion module are used to extract the feature information of the communication signal data;

[0042] S4. Train the signal modulation recognition model with the data set after data augmentation, and further optimize the signal modulation recognition model using a residual neural network to obtain the optimized signal modulation recognition model;

[0043] S5. Build a classifier for the optimized signal modulation recognition model.

[0044] In this embodiment, for step S1, first, the dataset needs to be divided into a training set and a test set, and the ratio of the training set to the test set is maintained at 7:3. After that, data standardization and alignment processing are also included, specifically as follows:

[0045] Perform data standardization on the data in the dataset:

[0046]

[0047] where x is the original data, μ is the average value of the dataset samples, σ is the standard deviation of the dataset samples, and x * is the data after standardization processing;

[0048] Perform sample alignment on the dataset: Translate the data in the dataset so that its data center of gravity g1 moves to nearby, where the calculation method of the data center of gravity g1 is as follows:

[0049]

[0050] In this embodiment, referring to Figure 2 , in order to increase the structural diversity of data samples and expand the dataset to avoid overfitting during model training, step S2 performs data augmentation on the dataset, specifically including using supervised single-sample data augmentation to process the data, transposing, merging, etc. on some signal data feature vectors to achieve the purpose of expanding the dataset, thereby avoiding overfitting during training and maximizing the use of useful information in the dataset.

[0051] In this embodiment, the recursive convolutional network performs layer-by-layer convolutional operations on multiple feature tensors in a recursive manner and fuses with a branch having a channel attention mechanism to integrate multi-scale feature information of the data. Its data processing process is specifically as follows:

[0052] 1) Perform a pooling layer operation on the signal data: Divide the input signal data feature tensor into several non-overlapping 2×2-sized square regions, and calculate the average value for the 4 values in each region.

[0053] Among them, the general calculation formula for the average value is:

[0054]

[0055] After the pooling layer operation, feature information can be further extracted, and the size of the feature map is reduced, which reduces the required computational amount and memory to a certain extent.

[0056] 2) Perform the pooling layer operation four times to obtain four identical feature tensors ci , then perform a convolutional operation on the recursive structure while introducing another branch. The branch processes the input using a channel attention mechanism and outputs it using a ReLU activation function and a batch normalization processing layer.

[0057] Specifically:

[0058] First, perform a convolutional operation on two of them. The convolutional kernel sizes are 1 and 3 respectively, without using padding. Pool the obtained two-dimensional larger feature tensors to make the sizes of the two feature tensors the same, and then perform an addition operation and output. Then, select one feature tensor to perform a convolutional operation with a convolutional kernel size of 5. After performing a convolutional operation with a convolutional kernel size of 1 on the previous output and then pooling, add the two to obtain a new feature tensor. And so on, execute the fourth tensor block. The simplified expression of the output C is as follows:

[0059] C = BN{Conv(BN{Conv(BN{Conv(c1, c2)}, c3)}, c4)}

[0060] Behind each convolutional layer, there is a ReLU activation function and a batch normalization processing layer connected in sequence, which better integrates the feature information.

[0061] 3) The recursive convolutional network also introduces another branch. This branch processes the input using a channel attention mechanism and also outputs it using a ReLU activation function and a batch normalization processing layer, enhancing the information correlation between channels of the feature tensor and enhancing the non-linear transformation ability of the model. After adding this output and the output of the recursive convolutional network, it is input to the subsequent self-attention mechanism layer for subsequent operations.

[0062] Specifically, through the self-attention mechanism, the model can better learn the correlation between different parts of multiple feature vectors. For each input vector a, multiply it by three coefficients w q , w k , w v to obtain three values q, k, and v:

[0063] q i = w q ·a i , k i = w k ·a i , v i = w v ·a i

[0064] Use the dot product method to calculate the correlation α between any two vectors i,j .

[0065] α i,j = qi ·k j

[0066] Let α i,j After passing the matrix A formed by through the ReLU activation function, calculate the output vector b corresponding to the input vector a with the matrix V i formed.

[0067]

[0068] In this embodiment, the data processing process of the multi-scale feature fusion module is specifically as follows:

[0069] First, perform batch normalization on the input, then perform convolution operations on each channel using three convolutions with kernel sizes of 3, 5, and 7 respectively, use zero-padding to ensure the same two-dimensional output size, then add them up and perform a 1×1 convolution operation to obtain a weight vector, and finally multiply it with the input to apply the weight to the feature tensor to obtain the output.

[0070] In this embodiment, the residual neural network includes a convolutional input layer and residual convolutional layers, and outputs through a fully connected layer. Among them, the convolutional input layer ensures the integrity of the data structure and information, which helps to extract feature information subsequently. The residual layer uses a convolutional layer with a kernel size generally of 3×3 to extract effective feature information from the enhanced signal. Each residual convolutional layer can also use the ReLU activation function to enhance the generalization ability of the model.

[0071] In this embodiment, the classifier includes a fully connected layer and a Softmax layer. Among them, the fully connected layer uses the method of matrix-vector multiplication to perform feature space transformation and integrate the local feature information extracted by the previous network layer, and multiplies the weight matrix with the input vector and adds a bias, and then classifies through the Softmax layer.

[0072] Specifically, the fully connected layer uses the method of matrix-vector multiplication to perform feature space transformation and integrate the local feature information extracted by the previous network layer, multiplies the weight matrix with the input vector and adds a bias, and then classifies through the Softmax layer. If the total number of targets contained in the training set is C, and the test sample X test The probability corresponding to the i-th type of target in the target set is expressed as:

[0073]

[0074] where x is the input of the fully connected layer, W T is the weight matrix, b is the bias, is the probability output by the -Softmax layer.

[0075] Through the maximum a posteriori probability, the test sample Xtest Classify into the maximum target probability c0:

[0076] c0 = argmax P(i│xtest)

[0077] The loss function of the classifier is generally designed as cross-entropy. The parameters are learned by calculating the gradient of the loss function with respect to the parameters using the training data, and the learned parameters are fixed when the model converges. The present invention adopts a cost function based on cross-entropy, which can be expressed as:

[0078]

[0079] where N represents the number of training samples in a batch, z(i) is used to represent the i-th training sample, and P(i│x train ) represents the probability that the training sample corresponds to the i-th target.

[0080] The advantages of the present invention are that it can fully extract the feature information of the signal, and can improve the calculation speed and recognition effect of the traditional signal modulation recognition method; further, the method progresses layer by layer in a recursive manner when processing the signal, and can effectively mine the deep feature information; furthermore, the method uses a channel attention mechanism to fuse the channel information, which is helpful for feature extraction; finally, the method performs additional preprocessing on the data in the pre-training stage to increase the diversity of communication signal samples. The pre-trained model will no longer be limited by a large number of training samples, which has practical significance in practical applications.

[0081] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements will not make the essence of the corresponding technical solutions deviate from the protection scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A signal modulation recognition method based on a recursive convolutional network and an attention mechanism, characterized in that It includes the following steps: S1. Sample each communication signal to generate a data set; S2. Perform data augmentation on the data set; S3. Establish a signal modulation recognition model, which includes a recursive convolutional network and a multi-scale feature fusion module arranged in sequence. The recursive convolutional network and the multi-scale feature fusion module are used to extract the feature information of communication signal data; The recursive convolutional network performs layer-by-layer convolutional operations on multiple feature tensors in a recursive manner and fuses with a branch having a channel attention mechanism to integrate multi-scale feature information of data; the multi-scale feature fusion module adopts convolutional operations with different convolutional kernel sizes and generates a weight vector to act on the input of the signal data; the data processing process of the recursive convolutional network is specifically as follows: perform a pooling layer operation on the signal data, divide the input signal data feature tensor into several non-overlapping 2 ×2 sized square regions, calculate the average value of the 4 values in each region; perform the pooling layer operation four times to obtain four identical feature tensors, and then perform a convolutional operation of a recursive structure, and at the same time introduce another branch, the branch uses the channel attention mechanism for the input, and uses the ReLU activation function and the batch normalization processing layer for output; the output of the recursive convolutional network and the output of the channel attention branch are feature-fused by an addition method; the data processing process of the multi-scale feature fusion module is specifically as follows. First, perform batch normalization processing on the input, and then use three convolutions with convolutional kernel sizes of 3, 5, and 7 respectively to perform convolutional operations on each channel, use the zero-padding method to ensure that the output two-dimensional sizes are the same, and then add them to perform 1 ×1 convolutional operation to obtain a weight vector, and finally multiply it with the input, and apply the weight to the feature tensor to obtain the output; S4. Use the data set after data augmentation to train the signal modulation recognition model, and further optimize the signal modulation recognition model using a residual neural network to obtain the optimized signal modulation recognition model; S5. Build a classifier for the optimized signal modulation recognition model.

2. The signal modulation recognition method based on a recursive convolutional network and an attention mechanism according to claim 1, wherein The data set includes a training set and a test set, and the data volumes of the training set and the test set are 7:

3.

3. The signal modulation recognition method based on a recursive convolutional network and an attention mechanism according to claim 1, wherein Step S1 further includes: Perform data standardization on the data in the data set; Among them, is the original data, is the average value of the samples in the said data set, is the standard deviation of the samples in the said data set, is the data after being standardized; Perform sample alignment on the dataset: Translate the data in the dataset so that the data center of gravity is moved to nearby, where the calculation method of the data center of gravity is as follows: 。 4. The signal modulation recognition method based on a recursive convolutional network and an attention mechanism according to claim 1, wherein Step S2 specifically includes: performing supervised single-sample data augmentation on the data in the data set, and performing transposition and merging operations on partial signal data feature vectors to expand the data set.

5. The signal modulation recognition method based on a recursive convolutional network and an attention mechanism according to claim 1, wherein The residual neural network includes a convolutional input layer and a residual convolutional layer, and outputs through a fully connected layer.

6. The signal modulation recognition method based on a recursive convolutional network and an attention mechanism according to claim 1, wherein The classifier includes a fully connected layer and a Softmax layer. Among them, the fully connected layer uses matrix-vector multiplication to perform feature space transformation and integrate the local feature information extracted by the previous network layer, and multiplies the weight matrix by the input vector and adds a bias, and then classifies through the Softmax layer.

Images