Signal modulation recognition method based on tree-structured perception fusion convolutional network
Through the combination of tree-perceptual fusion convolution network and feature compression module, the signal modulation recognition model is optimized, which solves the problems of cumbersome and poor recognition effect in the prior art, and achieves high recognition accuracy under low signal-to-noise ratio and robustness.
Patent Information
- Application Number
- CN202211661493.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-12-23
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2042-12-23
AI Technical Summary
The manual design features in the existing modulation recognition methods are cumbersome, and the recognition effect and robustness cannot be achieved, especially in the case of low signal-to-noise ratio, the recognition accuracy is low.
The tree-perceptual fusion convolution network and feature compression module are used to optimize the signal modulation recognition model through data augmentation and attention mechanism, eliminate artificial design features, and improve recognition effect and robustness.
High recognition accuracy in low signal-to-noise ratio is achieved, and the cumbersomeness of manual design features is eliminated, with good robustness and recognition effect.
Smart Images

Figure CN115987730B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of signal modulation recognition, and in particular to a signal modulation recognition method of a tree-like perception fusion convolutional network. Background Art
[0002] Communication satellites have the advantages of long communication distance, large capacity, good signal quality, high reliability and strong flexibility, and are widely used in the civilian field. In civilian communications, relevant government departments monitor civilian communication signals to achieve interference identification and electromagnetic spectrum management, while preventing illegal use and interference of wireless spectrum to ensure the normal conduct of legal communications. The modulation mode identification of communication signals is the premise and basis of signal demodulation, so it is of great significance to quickly and accurately identify the modulation mode of communication signals.
[0003] At present, modulation recognition methods are mainly divided into two categories: modulation recognition based on decision theory and modulation recognition based on machine learning. Modulation recognition based on decision theory is to manually extract some features of the signal and then realize modulation recognition through hypothesis testing; modulation recognition based on machine learning is to design a machine learning classifier and train it according to the characteristics of the signal classifier to achieve better modulation recognition and classification results. However, the current modulation method is cumbersome to manually design features and cannot achieve recognition effect and robustness. Summary of the invention
[0004] Technical issues solved
[0005] In view of the above-mentioned shortcomings of the prior art, the present invention provides a signal modulation recognition method based on a tree-like perception fusion convolutional network, which eliminates the tediousness of manually designed features and has good recognition effect and robustness.
[0006] Technical Solution
[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions:
[0008] The present invention provides a signal modulation recognition method based on a tree-like perception fusion convolutional network, comprising the following steps:
[0009] S1, collect communication signal data and generate data sets;
[0010] S2, performing data enhancement processing on the data set to achieve data set expansion;
[0011] S3, establishing a signal modulation recognition model, the signal modulation recognition model includes a tree-like perception fusion convolutional network and a feature compression module arranged in sequence, the tree-like perception fusion convolutional network is used to divide the input features into blocks and use a pooling operation to give the association information between the blocks, and then perform a tree-like structured convolution operation on each block, and each convolution layer is followed by a batch normalization processing layer; the feature compression module is used to further extract and compress the input feature information;
[0012] S4, using the data set after data enhancement to train the signal modulation recognition model, and using the attention mechanism to further optimize the signal modulation recognition model to obtain the optimized signal modulation recognition model;
[0013] S5. Building a classifier for the optimized signal modulation recognition model, and using it to perform signal modulation recognition.
[0014] Furthermore, the further extraction of input feature information and compression specifically includes: firstly dividing the input feature tensor into two blocks according to channel sampling, taking one block for pooling operation, then extracting feature information through convolution layers respectively, and finally fusing the output.
[0015] Furthermore, the data set includes a training set and a test set, and the data volume of the training set and the test set is maintained at 7:3.
[0016] Furthermore, step S1 also includes performing intensity normalization and center of gravity alignment processing on the data of the data set.
[0017] Furthermore, the intensity normalization is specifically as follows: the original data X raw It is expressed as:
[0018] X raw =[x 1 ,x 2 ,…,x L1 ],
[0019] Among them, L 1 represents the total number of distance units contained in the data of the data set, and the data x after intensity normalization is:
[0020]
[0021] The gravity center alignment process is specifically as follows: translating the data of the data set to make its data gravity center g 1 Move to Nearby, where the data center g 1 The calculation method is as follows:
[0022]
[0023] Furthermore, the specific operation of the tree-like perception fusion convolutional network is:
[0024] Split the input feature tensor so that different partitions can share parameters. Split an N×C×14×14 input into 4N×C×7×7, where N represents the number of data batches and each 7×7 block is considered as a partition.
[0025] Each channel of each partition is processed by global average pooling to become N×C×2×2, and then processed by the BN layer. Then the feature tensor is flattened into a one-dimensional vector and input into a fully connected layer. After the output, the value on the corresponding channel partition is added to all the values of the original partition, and finally passed to the subsequent layer as the output;
[0026] Perform convolution operations with a convolution kernel size of 1×1 on two of the four partitions to obtain the block C of the feature tensor 1 , and the other two are convolution operations with 3×3 convolution kernel, patch size 1 and 5×5 convolution kernel, patch size 2, respectively, to obtain blocks C 2 and block C 3 ;
[0027] The block C 1 and the block C 2 Add together, the block C 1 and the block C 3 Add them together, and get two new feature tensors by taking the average of the corresponding position values as the result. Then perform a convolution operation with a kernel size of 7×7 and a patch size of 3 on the two tensors, and add the obtained results together after passing through the ReLU activation function and batch normalization processing layer to obtain the output result C 4 ;
[0028] The output result C will be 4 The dimensionality is reduced to a one-dimensional feature vector, and the feature vector is input into two consecutive fully connected layers for processing. In addition, the two fully connected layers are followed by a ReLU activation function operation and then the output is merged and the output is combined with the output result C 4 The final output C is obtained by taking the average value of the same position and adding them together. out .
[0029] Furthermore, the specific operation of the feature compression module is:
[0030] First, the input feature tensor is sampled according to the channel, each channel is sampled once, and the feature tensor is divided into two tensors, and then an average pooling operation is performed on one of them. The average pooling operation takes a stride of 1, a window size of 2, and a sampling zero-filling method;
[0031] Performing ordinary convolution operations on the pooled feature tensor blocks and the unprocessed feature tensor blocks respectively, wherein the convolution operation takes the convolution kernel size of 3, the stride of 1, does not use zero padding, and each convolution is followed by a ReLU activation function and a batch normalization processing layer;
[0032] The two feature tensor blocks obtained are rearranged in the sampling order and restored to the original arrangement order for output.
[0033] Furthermore, the attention mechanism specifically includes:
[0034] Through the feature matrix F, historical attention probability and β t To calculate the current attention weight α t , and then get the output vector c t :
[0035]
[0036]
[0037] Among them, α l represents the attention weight at time step l, which is calculated as follows:
[0038] α t = sinh(U f F+W s N s +W l N l )
[0039] Among them, U f is a 1×1 convolution operation, W s , W l is the trainable weight matrix, N s , N l are the products of two convolution kernel parameters and historical attention probabilities, and t represents time.
[0040] Beneficial Effects
[0041] The present invention designs a signal modulation recognition method based on a tree-like perception fusion convolutional network. The signal modulation recognition method based on a tree-like perception fusion convolutional network and feature compression is used to combine feature extraction and classifier in one framework, thereby eliminating the tediousness of manually designed features and having good recognition effect and robustness. In addition, the method completes signal modulation recognition through the output layer. Under the condition of low signal-to-noise ratio, the method has higher recognition accuracy than traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0043] Figure 1 A schematic diagram of the steps of a signal modulation recognition method based on a tree-like perception fusion convolutional network provided by an embodiment of the present invention;
[0044] Figure 2 A model framework diagram of a signal modulation recognition method based on a tree-structured perception fusion convolutional network provided in one embodiment of the present invention. DETAILED DESCRIPTION
[0045] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0046] See also Figure 1 and Figure 2 An embodiment of the present invention provides a signal modulation recognition method based on a tree-like perception fusion convolutional network, comprising the following steps:
[0047] S1, collect communication signal data and generate data sets;
[0048] S2, performing data enhancement processing on the data set to achieve data set expansion;
[0049] S3, establishing a signal modulation recognition model, the signal modulation recognition model includes a tree-like perception fusion convolutional network and a feature compression module arranged in sequence, the tree-like perception fusion convolutional network is used to divide the input features into blocks and use a pooling operation to give the association information between the blocks, and then perform a tree-like structured convolution operation on each block, and each convolution layer is followed by a batch normalization processing layer; the feature compression module is used to further extract and compress the input feature information;
[0050] S4, using the data set after data enhancement to train the signal modulation recognition model, and using the attention mechanism to further optimize the signal modulation recognition model to obtain the optimized signal modulation recognition model;
[0051] S5. Building a classifier for the optimized signal modulation recognition model, and using it to perform signal modulation recognition.
[0052] In this embodiment, see Figure 2 For step S1, the data set includes a training set and a test set, and the data volume of the training set and the test set is maintained at 7:3. Step S1 also includes performing intensity normalization and center of gravity alignment processing on the data of the data set.
[0053] The intensity normalization is specifically as follows: raw It is expressed as:
[0054] X raw =[x 1 ,x 2 ,…,x L1 ],
[0055] Among them, L 1 represents the total number of distance units contained in the data of the data set, and the data x after intensity normalization is:
[0056]
[0057] The gravity center alignment process is specifically as follows: translating the data of the data set to make its data gravity center g 1 Move to Nearby, where the data center g 1 The calculation method is as follows:
[0058]
[0059] In this embodiment, step S2 is to perform data enhancement, specifically: two signal samples are randomly selected from the data set by using a sample pairing method, and after transposing them, the two samples are superimposed into a new sample by taking the average value, and the label of the new sample is any one of the two samples, so as to achieve the purpose of expanding the data set.
[0060] In this embodiment, see Figure 2 , the further extraction of input feature information and compression specifically includes: firstly, the input feature tensor is divided into two blocks according to channel sampling, and then one block is taken for pooling operation, and then the feature information is extracted through convolution layers respectively, and finally the output is fused.
[0061] In this embodiment, see Figure 2 , the specific operation of the tree-like perception fusion convolutional network is as follows:
[0062] 1) The tree-like perception fusion convolutional network first separates the feature tensor so that different partitions can share parameters. An N×C×14×14 input can be divided into 4N×C×7×7, where N is the number of data batches, and each 7×7 block is regarded as a partition. The split breaks the correlation between different partitions of the same channel. In order to increase the correlation in each partition, each channel of each partition is first processed by global average pooling to become N×C×2×2, and then processed by the BN layer to increase the nonlinear transformation ability of the feature tensor. Then, the squeeze() operation is used to flatten the tensor into a one-dimensional vector input into a fully connected layer. After the output, it is reshaped to add the values on the corresponding channel partition to all the values of the original partition, and finally passed to the subsequent layers as output.
[0063] 2) The multi-scale perception module uses multiple convolution kernels of different sizes to perform convolution operations on the grouped feature tensors, and each convolution layer is followed by a batch normalization processing layer. Two of the four partitions are convolved with a convolution kernel size of 1×1 to obtain feature tensor blocks C 1 , and the other two convolution operations using 3×3 convolution kernel, patch size 1 and 5×5 convolution kernel, patch size 2, respectively, get C 2 and C 3 The purpose of introducing the patch operation is to keep the two-dimensional size of the result of the feature tensor block after the convolution operation unchanged.
[0064] 3) Divide the feature tensor of the same size obtained in the previous step into blocks C 1 and C 2 Add, C 1 and C 3 Add them together, and get two new feature tensors by taking the average of the corresponding position values as the result. Then perform a convolution operation with a kernel size of 7×7 on the two tensors. Also, in order to maintain the two-dimensional size unchanged, a patch size of 3 is introduced. The results are added together and then passed through the ReLU activation function and batch normalization processing layer, and then spliced to obtain the output result C 4 .
[0065] 4) Use the squeeze() operation to reduce the dimension of the feature tensor after block processing to a one-dimensional feature vector, and input the vector into two consecutive fully connected layers for processing. Both fully connected layers are followed by a ReLU activation function operation, and their output is summed with C 4 The final output C is obtained by taking the average value of the same position and adding them together. out .
[0066] In this embodiment, see Figure 2 , the specific operation of the feature compression module is as follows:
[0067] First, the input feature tensor is sampled according to the channel, each channel is sampled once, and the feature tensor is divided into two tensors, and then an average pooling operation is performed on one of them. The average pooling operation takes a stride of 1, a window size of 2, and a sampling zero-filling method;
[0068] Performing ordinary convolution operations on the pooled feature tensor blocks and the unprocessed feature tensor blocks respectively, wherein the convolution operation takes the convolution kernel size of 3, the stride of 1, does not use zero padding, and each convolution is followed by a ReLU activation function and a batch normalization processing layer;
[0069] The two feature tensor blocks obtained are rearranged in the sampling order and restored to the original arrangement order for output.
[0070] The above output is generally output to the attention mechanism module to achieve model optimization.
[0071] Specifically, the attention mechanism includes:
[0072] Through the feature matrix F, historical attention probability and β t To calculate the current attention weight α t , and then get the output vector c t :
[0073]
[0074]
[0075] Among them, α l represents the attention weight at time step l, which is calculated as follows:
[0076] α t = sinh(U f F+W s N s +W l N l )
[0077] Among them, U f is a 1×1 convolution operation, W s , W l is the trainable weight matrix, N s , N l are the products of two convolution kernel parameters and historical attention probabilities, and t represents time.
[0078] In this embodiment, the classifier is a linear layer, which includes a fully connected layer and a Softmax layer. The fully connected layer uses matrix-vector multiplication to transform the feature space and integrate the local feature information extracted by the previous network layer. The weight matrix is multiplied by the input vector and then the bias is added and then classified by the Softmax layer. If the total number of targets included in the training set is C, the test sample X test The probability corresponding to the i-th target in the target set is expressed as:
[0079]
[0080] Where x is the input of the fully connected layer, W T is the weight matrix, b is the bias, is the probability of the Softmax layer output.
[0081] Finally, the test sample X is converted to test Classification to the maximum target probability c 0 middle:
[0082] c 0 =argmaxP(i│x test )
[0083] In this embodiment, the loss function of the classifier is designed to be cross entropy, so as to learn parameters by calculating the gradient of the loss function with respect to the parameters using training data, and fix the learned parameters when the model converges.
[0084] The present invention adopts a cost function based on cross entropy, which can be expressed as:
[0085]
[0086] 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.
[0087] The advantage of the present invention is that it combines feature extraction and classifier in one framework by using a signal modulation recognition method based on a tree-like perception fusion convolutional network and feature compression, eliminating the tediousness of manually designing features and having good recognition effect and robustness; in addition, the method completes signal modulation recognition through an output layer, and under low signal-to-noise ratio conditions, the method has higher recognition accuracy than traditional methods.
[0088] 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 aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements will not cause the essence of the corresponding technical solutions to deviate from the protection scope of the technical solutions of the embodiments of the present invention.
Claims
1. A signal modulation recognition method based on tree-like perception fusion convolutional network, characterized in that: The steps include: S1, collect communication signal data and generate data sets; S2, performing data enhancement processing on the data set to achieve data set expansion; S3, establishing a signal modulation recognition model, the signal modulation recognition model includes a tree-like perception fusion convolutional network and a feature compression module arranged in sequence, the tree-like perception fusion convolutional network is used to divide the input features into blocks and use a pooling operation to give the association information between the blocks, and then perform a tree-like structured convolution operation on each block, and each convolution layer is followed by a batch normalization processing layer; The feature compression module is used to further extract and compress input feature information; S4, using the data set after data enhancement to train the signal modulation recognition model, and using the attention mechanism to further optimize the signal modulation recognition model to obtain the optimized signal modulation recognition model; S5. Building a classifier for the optimized signal modulation recognition model, and using it to perform signal modulation recognition; The specific operation of the tree-like perception fusion convolutional network is: Split the input feature tensor so that different partitions can share parameters. Split an N×C×14×14 input into 4N×C×7×7, where N represents the number of data batches and each 7×7 block is considered as a partition. Each channel of each partition is processed by global average pooling to become N×C×2×2, and then processed by the BN layer. Then the feature tensor is flattened into a one-dimensional vector and input into a fully connected layer. After the output, the value on the corresponding channel partition is added to all the values of the original partition, and finally passed to the subsequent layer as the output; Perform convolution operations with a convolution kernel size of 1×1 on two of the four partitions to obtain block C1 of the feature tensor, and perform convolution operations with a convolution kernel size of 3×3 and a patch size of 1 and a convolution kernel size of 5×5 and a patch size of 2 on the other two to obtain blocks C2 and C3 respectively; Add the block C1 and the block C2, and add the block C1 and the block C3, and obtain two new feature tensors by taking the average of the corresponding position values as the result, and then perform a convolution operation with a kernel size of 7×7 and a patch size of 3 on the two tensors, and then add the obtained results and perform a splicing operation through a ReLU activation function and a batch normalization processing layer to obtain an output result C4; The output result C4 is reduced to a one-dimensional feature vector, and the feature vector is input into two consecutive fully connected layers for processing. The two fully connected layers are followed by a ReLU activation function operation and then the outputs are merged. The outputs are added to the output result C4 by taking the average value of the same position to obtain the final output C out .
2. The signal modulation recognition method based on tree-structured perception fusion convolutional network according to claim 1 is characterized in that: The further extraction of input feature information and compression specifically includes: firstly, dividing the input feature tensor into two blocks according to channel sampling, taking one block for pooling operation, then respectively extracting feature information through convolution layers, and finally fusing the output.
3. The signal modulation recognition method based on tree-structured perception fusion convolutional network according to claim 2 is characterized in that: The data set includes a training set and a test set, and the data volume of the training set and the test set is maintained at 7:
3.
4. The signal modulation recognition method based on tree-structured perception fusion convolutional network according to claim 3 is characterized in that: Step S1 also includes performing intensity normalization and center of gravity alignment processing on the data of the data set.
5. The signal modulation recognition method based on tree-structured perception fusion convolutional network according to claim 4 is characterized in that: The intensity normalization is specifically as follows: raw It is expressed as: Wherein, L1 represents the total number of distance units contained in the data of the data set, and the data x after intensity normalization is: The gravity center alignment process is specifically as follows: translating the data of the data set so that the data gravity center g1 moves to Nearby, wherein the calculation method of the data center g1 is as follows:
6. The signal modulation recognition method based on tree-structured perception fusion convolutional network according to claim 1 is characterized in that: The specific operations of the feature compression module are: First, the input feature tensor is sampled according to the channel, each channel is sampled once, and the feature tensor is divided into two tensors, and then an average pooling operation is performed on one of them. The average pooling operation takes a stride of 1, a window size of 2, and a sampling zero-filling method; Performing ordinary convolution operations on the pooled feature tensor blocks and the unprocessed feature tensor blocks respectively, wherein the convolution operation takes the convolution kernel size of 3, the stride of 1, does not use zero padding, and each convolution is followed by a ReLU activation function and a batch normalization processing layer; The two feature tensor blocks obtained are rearranged in the sampling order and restored to the original arrangement order for output.
7. The signal modulation recognition method based on tree-structured perception fusion convolutional network according to claim 6 is characterized in that: The attention mechanism specifically includes: Through the feature matrix F, historical attention probability and β t To calculate the current attention weight α t , and then get the output vector c t : Among them, α l represents the attention weight at time step l, which is calculated as follows: α t =birth(U) f F+W s N s +W l N l ) Among them, U f is a 1×1 convolution operation, W s , W l is the trainable weight matrix, N s , N l are the products of two convolution kernel parameters and historical attention probabilities, and t represents time.
