Small sample interference signal classification and identification method based on local-global fusion network
Through the local-global fusion network, combined with multi-scale expansion convolution, multi-head self-attention and multi-convolution kernel high-efficiency channel attention module, the problem of insufficient feature extraction in small sample interference signal recognition is solved, and efficient recognition under low data volume conditions is achieved.
Patent Information
- Application Number
- CN202510354229.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-25
- Publication Date
- 2025-07-11
AI Technical Summary
In the case of small samples, it is difficult for the prior art to effectively extract the characteristics of the interference signal, resulting in inaccurate identification results. Traditional methods require a large amount of sample data, which makes data acquisition difficult and poor channel conditions affect the model training effect.
The local-global fusion network is adopted to extract the local and global features of the interference signal through multi-scale expansion of the convolution module, the multi-head self-attention module and the multi-convolution kernel efficient channel attention module, enhance the information interaction between channels, and build a local-global fusion network model for classification identification.
It improves the accuracy of interference signal recognition of small samples, reduces the requirements for data volume, solves the problem of insufficient feature extraction in small samples, and realizes efficient identification in complex communication environments.
Smart Images
Figure CN120296494A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of communication technologies, and specifically, relates to a method for classifying and identifying small-sample interference signals based on a local-global fusion network. Background Art
[0002] Currently, there are mainly two research approaches for the classification and identification of interference signals: one is the classification and identification of interference signals based on multi-scale fuzzy entropy (MFE); the other is the classification and identification of interference signals based on AFL. AFL (American FuzzyLop) is one of the most popular fuzz testing tools currently, released by Google security engineer Michal Zalewski.
[0003] The classification and identification of interference signals based on MFE mainly involve two aspects of research, namely feature extraction and classification and identification. In terms of feature extraction, the feature parameters extracted manually generally have clear physical meanings, that is, the physical meanings of the extracted feature parameters are interpretable. In addition, the feature parameters extracted manually have high discrimination, can reflect the differences of different interference signal types, and are convenient for the classifier to accurately classify. The feature extraction of interference signals is generally carried out in the time domain, frequency domain, time-frequency domain, and variation domain, and its purpose is to convert the original communication interference signal into a set of numerical representations that can reflect the characteristics of the interference signal. Classification and identification are generally carried out under the premise of MFE, and the extracted feature parameters are input into the classifier to achieve classification and identification. The classification and identification algorithms used based on MFE are generally traditional machine learning algorithms, such as support vector machine (SVM) classifiers, decision tree (DT) classifiers, Bayesian classifiers, artificial neural networks (ANN), BP neural networks (BPNN), deep neural networks (DNN), etc. Among them, the classifiers composed of SVM, DT, and Bayesian algorithms can provide the interpretability of classification decisions, and can understand the classification results through tree diagrams or probability distributions, etc. While ANN, BPNN, and DNN are classification algorithms based on the neural network principle, which can effectively process high-dimensional data containing a large number of features, but it is difficult to provide the interpretability of classification decisions. Therefore, the above methods have unsatisfactory feature extraction effects in the case of small samples.
[0004] AFL-based communication interference classification and recognition refers to automatically learning the features of interference signals through a convolutional neural network (CNN) and completing classification and recognition without manual extraction of feature parameters, reducing the huge workload required by MFE. The neural networks currently used include convolutional neural networks (CNN), residual networks (ResNet), long short-term memory networks (LSTM), etc. When processing interference signals, CNN and ResNet often need to convert the interference signals into two-dimensional data, and then extract features and classify the interference signals. The communication interference signals actually simulated are often one-dimensional long sequences, and the signal sequences are correlated. When preprocessing the interference signal sequences, it will cause partial information loss of the signals and certain damage to the temporal correlation of the signals. In the case of small samples, the above disadvantages are more obvious.
[0005] Generally speaking, the modern communication environment is complex and severe, and the challenges faced by communication anti-interference are increasing day by day. Due to the high frequency and fast change of new interference signals, the traditional communication interference signal recognition methods based on a large number of samples face two urgent problems: First, the complex communication environment makes it extremely difficult to obtain sufficient and representative samples, restricting the development of data-driven methods such as deep learning; Second, the channel conditions in the actual communication environment are harsh, and it is difficult to train the model effectively only relying on a small number of collected samples, resulting in limitations on the performance and generalization ability of the model. Summary of the Invention
[0006] Aiming at the problems existing in the prior art, the object of the present invention is to propose a small-sample interference signal classification and recognition method based on a local-global fusion network, jointly extract global features and local features, increase the weight of important features, improve the recognition accuracy of interference signals in small-sample scenarios, solve the problem that in the case of small samples, the existing network fails to extract features sufficiently, resulting in incorrect recognition results, has a low requirement for the amount of data, and breaks through the limitation that the existing small-sample interference signal classification and recognition methods all first expand the data and then perform classification and recognition.
[0007] The technical solution adopted by the present invention to achieve the above object is: to propose a small-sample interference signal classification and recognition method based on a local-global fusion network, including the following steps:
[0008] Step 1: Construct a data set
[0009] Collect interference signals, assign different tags according to their categories, extract the in-phase component I and quadrature component Q of the interference signals. To conduct a more comprehensive analysis of the signals, take the modulus of the signals to construct a dataset, and then perform normalization preprocessing on the data in the constructed dataset to obtain the original interference signal dataset. Randomly select a set proportion of the data from the original interference signal dataset as the small-sample interference signal dataset; preferably, take 8% of the sample quantity of the original interference signal dataset as the small-sample interference signal dataset; and divide the small-sample interference signal dataset into a training set and a test set;
[0010] The normalization process is shown in the following formula:
[0011]
[0012] where x′ is the data after normalization processing; x is the data in the dataset, min(x) is the maximum value of the data in the dataset, and max(x) is the minimum value of the data in the dataset.
[0013] Step 2: Construct a local-global fusion network model, and use the training set and test set to iteratively train the constructed local-global fusion network model to obtain a trained local-global fusion network model; the local-global fusion network model includes a multi-scale dilated convolution module, a multi-head self-attention module, and a multi-kernel efficient channel attention (MKP-ECA) module. The multi-scale dilated convolution module and the multi-head self-attention module are arranged in parallel to extract local features and global features respectively. Subsequently, stack the output features of the multi-scale dilated convolution module and the multi-head self-attention module on the channels to obtain an output feature map with local and global features where and are the output features of the multi-scale dilated convolution module and the multi-head self-attention module respectively, C represents the number of channels of the feature map, and T represents the time dimension of the feature map; finally, the output feature map with local and global features is input into the multi-kernel efficient channel attention module. Let the input feature vector be In the multi-kernel efficient channel attention module, first, use a global pooling layer to compress the input features into an aggregated channel to obtain a representative value for each channel; secondly, use three convolutional weights with different kernel sizes to scale and weight the aggregated channel to enhance the information interaction between channels; activate with the Sigmoid function to obtain a scaled channel weight within the range of 0-1 Multiply the scaled channel weight element-wise with the input feature vector to obtain the output feature vector The calculation process of the multi-kernel efficient channel attention module is expressed as:
[0014]
[0015] where w k1 , w k2 and w k3 represent the convolution weights with kernel sizes of 3, 5, and 7 respectively, σ is the SoftMax function; GAP represents the global pooling layer;
[0016] Step 3: Input the small-sample interference signal to be recognized into the trained local-global fusion network model for classification and recognition to obtain the recognition result.
[0017] Furthermore, the multi-scale dilated convolution module is configured as follows:
[0018] It adopts three parallel branch structures; subsequently, the output elements of the three parallel branches are stacked using the Concat function on the channel dimension to form an overall feature map; and a convolutional layer with a kernel of 3 is used to solve the aliasing problem caused by the concatenation operator; this process is expressed as follows:
[0019]
[0020] where represents the input feature vector; represents the output feature of the multi-scale dilated convolution module, C represents the number of channels of the feature map, and T represents the input time dimension; represents the dilated convolution weight with a dilation rate of r, and its kernel size is 5; represents the convolution weight with a kernel of 3, H r represents the output features of convolutions with different dilation rates.
[0021] Preferably, the dilation rates adopted by the three parallel branches are 1, 2, and 5 respectively.
[0022] Furthermore, the multi-head self-attention module is configured as follows:
[0023] Let the input feature vector be C represents the number of channels of the feature map, and T represents the input time dimension; first, perform three parallel linear transformations for feature mapping to respectively obtain the query matrix φ h , the key matrix τ h and the value matrix ν h , This process is expressed as follows:
[0024]
[0025] where, and respectively represent the linear mapping weights of the query matrix, the key matrix, and the value matrix; C h represents the feature dimension, and the query matrix, the key matrix, and the value matrix have the same feature dimension; h represents the number of heads of the multi-head self-attention module, where C = h × C h ;
[0026] Subsequently, the transpose of the query matrix φ h and the key matrix τ h is subjected to a dot product operation to obtain the similarity value between the feature maps; then, scaling and normalization operations are performed on it to obtain the attention weights, and these attention weights are further multiplied by the value matrix ν h to obtain the weighted features The process is as follows:
[0027]
[0028] where σ is the SoftMax function, representing the normalization operation, and tr represents the matrix transpose;
[0029] Stack the weighted features output by all heads on the channels and perform feature mapping to obtain the output features which is shown as follows:
[0030]
[0031] where represents the output features of the multi-head self-attention module; represents the convolution weight with a kernel of 3.
[0032] Through the above design, the present invention can bring the following beneficial effects: The present invention proposes a method for classifying and identifying small-sample interference signals based on a local-global fusion network, collects interference signals, assigns different labels according to their categories to obtain the original interference signal dataset, and randomly selects a set proportion of data from the original interference signal dataset as the small-sample interference signal dataset; relies on the multi-scale dilation convolution module to extract the local fine features of the data, uses the multi-head self-attention module to learn the long-term and global features of the signals, designs a multi-convolution kernel efficient channel attention module to capture the dependencies between channels, enhances the information interaction between channels, and improves the attention to important channels; constructs a local-global fusion network model by combining the above modules to classify and identify small-sample interference signals. The present invention solves the problem that in the case of small samples, the existing networks extract features insufficiently, resulting in incorrect recognition results, has a low requirement for the amount of data, and breaks through the limitations of the existing methods for classifying and identifying small-sample interference signals. Brief Description of the Drawings
[0033] The accompanying drawings described herein are used to provide a further understanding of the present invention and form a part of this application. The illustrative embodiments of the present invention and their descriptions are used to understand the present invention and do not constitute an improper limitation of the present invention. In the drawings:
[0034] Figure 1 It is a flow chart of a small-sample interference signal classification and recognition method based on a local-global fusion network;
[0035] Figure 2 It is a schematic diagram of the structure framework of a local-global fusion network;
[0036] Figure 3 It is a functional structure diagram of a multi-scale dilated convolution module;
[0037] Figure 4 It is a functional structure diagram of a multi-head self-attention module;
[0038] Figure 5 It is a functional structure diagram of a multi-convolution kernel efficient channel attention module;
[0039] Figure 6 It is a flow chart of using three serially deployed local-global fusion networks in this embodiment to classify and recognize small-sample interference signals;
[0040] Figure 7 It is a result graph of the recognition rate of each of the 5 types of jamming signals in the embodiment of the present invention. Detailed implementation manners
[0041] To make the objectives, features, and advantages of the present invention more obvious and understandable, the technical solutions in the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention and the accompanying drawings. Obviously, the present invention is not limited by the following embodiments, and the specific implementation manners can be determined according to the technical solutions of the present invention and the actual situation. To avoid obscuring the essence of the present invention, well-known methods, processes, and procedures are not described in detail. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.
[0042] The present invention proposes a small-sample interference signal classification and recognition method based on a local-global fusion network, Figure 1 shows a flow chart of a small-sample interference signal classification and recognition method based on a local-global fusion network; the method includes: collecting interference signals, assigning different labels according to their categories, extracting the in-phase component I and the quadrature component Q of the interference signals, and taking the modulus of the signals to construct a data set for a more comprehensive analysis of the signals, and then performing normalization preprocessing. After normalization preprocessing, the original interference signal data set is obtained. The normalization process is shown in the following formula:
[0043]
[0044] Among them, x' is the data after normalization; x is the data in the dataset, min(x) is the maximum value of the data in the dataset, and max(x) is the minimum value of the data in the dataset.
[0045] If the data is not normalized and the original data is directly used, then the differences in the eigenvalues of the data are relatively large, which will cause large fluctuations during gradient descent and too long training time. Normalizing the data of the suppression interference signal is mainly to limit the data within the range of [0, 1] to eliminate the adverse effects brought by the outliers in the dataset. After normalization, the data in the dataset fluctuates less, and the optimization process of the optimal solution during gradient descent is smoother, making it easier to converge to the optimal solution, accelerating the convergence during the training process of the network model, and at the same time helping to improve the recognition accuracy of the network.
[0046] Randomly select a set proportion of data from the original interference signal dataset as the small-sample interference signal dataset, and divide the small-sample interference signal dataset into a training set and a test set;
[0047] Construct a local-global fusion network model, and use the training set and the test set to iteratively train the constructed local-global fusion network model to obtain a trained local-global fusion network model; the local-global fusion network model includes a multi-scale dilated convolution module, a multi-head self-attention module, and a multi-kernel efficient channel attention (MKP-ECA) module; and use the training set and the test set to iteratively train the constructed local-global fusion network model to obtain a trained local-global fusion network model; input the small-sample interference signal to be recognized into the trained local-global fusion network model for classification and recognition to obtain the recognition result.
[0048] It should be noted that the modulus of a signal refers to the amplitude of the signal in the complex plane and is usually used to describe the strength of the signal. For a complex signal, the modulus of the signal is where I is the in-phase component (real part) and Q is the quadrature component (imaginary part).
[0049] In the case of small samples, the traditional method has the problem of insufficient feature extraction. Aiming at the problem that it is difficult to effectively train the model in the small-sample scenario, the present invention extracts the non-linear features of the interference signal, establishes the correlation between the global and local signal features, and constructs a new type of local-global fusion network (GLFE Module) guided by the multi-head self-attention mechanism and multi-scale dilated convolution, thus avoiding this problem. At the same time, it ensures that the network can converge quickly during the training process and avoid falling into the local optimal solution. The fused global and local features contain a large number of redundant features. In order to capture the representative dependencies between channels and improve the attention to important channels, the present invention constructs an MKP-ECA module. The recognition accuracy of the interference signal is improved under the condition of insufficient sample size. Figure 2 Shows the schematic diagram of the local-global fusion network structure framework; Figure 3 Shows the functional structure diagram of the multi-scale dilated convolution module; Figure 4 Shows the functional structure diagram of the multi-head self-attention module; Figure 5 Shows the functional structure diagram of the multi-convolution-kernel efficient channel attention module, Figure 5 The Add instruction is widely used in computers. The Add instruction is used to perform addition operations; in this embodiment, the small-sample interference signal classification and recognition method based on the local-global fusion network includes:
[0050] 1. Construct a data set:
[0051] In order to be closer to the real complex electromagnetic environment, Gaussian white noise is added, and MATLAB software is used to simulate 5 types of jamming signals. The 5 types of jamming signals include single-tone jamming signals, multi-tone jamming signals, partial-band jamming signals, noise frequency modulation jamming signals, and noise amplitude modulation jamming signals.
[0052] The simulation parameters are as follows: the center frequencies of the 5 types of jamming signals are set to 30 MHz, the sampling frequency is set to 200 MHz, the bandwidth of the partial-band jamming signal is set to 40 MHz, and the modulation degree of the noise amplitude modulation jamming signal is set to 0.5. The interference-to-noise ratio (INR) is controlled between -20 dB and 0 dB, and data is collected every 2 dB. In each case of the interference-to-noise ratio, 1000 groups of data of the same type of signal are collected, and each group of data contains 3599 sampling points. Through the collection of simulation data, the size of the original interference signal data set constructed is 5000×3599. Since the sampling data of the interference signal is in complex form, the in-phase component I and the quadrature component Q of the interference signal are extracted. In order to analyze the signal more comprehensively, the modulus of the signal is taken to construct a data set, and then normalized preprocessing is performed to generate the original interference signal data set.
[0053] Add tags from 0 to 4 in the order of 5 kinds of jamming signals of suppression type. 8% of the sample number of the original jamming signal dataset is used as the small-sample jamming signal dataset, and the collected small-sample jamming signal dataset is randomly divided into a training set and a test set according to the ratio of 7:3 for training the network model.
[0054] 2. Model construction:
[0055] For the one-dimensional time-domain dataset of the generated 5 kinds of jamming signals of suppression type, a network for the classification and recognition method of small-sample jamming signals with a local-global fusion network is designed for training.
[0056] First, the data in the training set of the small-sample jamming signal dataset are input in parallel into a multi-scale dilated convolution module and a multi-head self-attention module to extract local and global features respectively, so as to solve the problem of insufficient feature extraction caused by insufficient data in the case of small samples. Subsequently, the outputs of the multi-scale dilated convolution module and the multi-head self-attention module are stacked on the channel respectively, and an output feature map with local and global features can be obtained. Finally, it is put into the MKP-ECA module to capture the representative dependencies between channels and improve the attention to important channels.
[0057] The multi-scale dilated convolution module uses convolutions with multiple different dilation rates to extract features from multiple local receptive fields, fuse the information of different channels, and thus capture more features hidden in the input information. Considering that if the dilation rate is too large, the local correlation of the information will be lost, the present invention adopts a zigzag structure to set the dilation rates of the three parallel branches to 1, 2, and 5 respectively; subsequently, the output features are stacked on the channel using the existing Concat function to form an overall feature map; and a convolution layer with a kernel of 3 is used to solve the aliasing problem caused by the concatenation operator, which can ensure the stability of the features. This process is shown as follows:
[0058]
[0059] where represents the input feature vector; represents the output feature of the multi-scale dilated convolution module, C represents the number of channels of the feature map, and T represents the input time dimension; represents the dilation convolution weight with a dilation rate of r, and its convolution kernel size is 5; represents the convolution weight with a kernel of 3, H r represents the output features of convolutions with different dilation rates. The Concat function is a function widely used in various programming and database environments, mainly used to concatenate two or more strings to form a new single string, which belongs to the prior art and will not be elaborated here in detail.
[0060] The multi-head self-attention module, as the core component of Transformer, can adaptively allocate attention weights from feature data to obtain long-term, non-local dependencies and further focus on the representative information contained in multi-scale features. Let the input feature vector be C represents the number of channels of the feature map, and T represents the input time dimension. First, through three parallel linear transformations (Linear) for feature mapping, the query matrix φ h , the key matrix τ h and the value matrix ν h are obtained respectively. This process is expressed as follows:
[0061]
[0062] Among them, and represent the linear mapping weights of the query matrix, key matrix, and value matrix respectively; C h represents the feature dimension, and the query matrix, key matrix, and value matrix have the same feature dimension; h represents the number of heads of the multi-head self-attention module, where C = h × C h .
[0063] Subsequently, the dot product operation is performed on the query matrix φ h and the transpose of the key matrix τ h to obtain the similarity value between feature maps. Then, scaling and normalization operations are performed on it to obtain the attention weights, and these attention weights are further multiplied by the value matrix ν h to obtain the weighted features The process is expressed as follows:
[0064]
[0065] where σ is the SoftMax function, representing the normalization operation, and tr represents the matrix transpose.
[0066] Stack the weighted features output by all heads on the channel and perform feature mapping to obtain the output feature It is expressed as follows:
[0067]
[0068] Among them represents the output feature of the multi-head self-attention module; represents the convolution weight with a kernel of 3;
[0069] Stack the outputs of the multi-scale dilated convolution module and the multi-head self-attention module on the channel using the existing Concat function to obtain the output feature map with local and global features The process is expressed as:
[0070]
[0071] The above method solves the problem of insufficient feature extraction of small-sample data. However, the fused global and local features contain a large number of redundant features. To capture the representative dependence relationships between channels and improve the attention to important channels, the present invention constructs an MKP-ECA module. Let the input feature vector be The global pooling layer is used to compress the input features into aggregated channels to obtain the representative value of each channel. Secondly, convolutional weights with three different kernel sizes are used to scale and weight the aggregated channels to enhance the information interaction between channels. The Sigmoid function is used for activation to obtain the scaled channel weights in the range of 0-1 The scaled channel weights and the input feature vector are multiplied element by element to obtain the output feature vector The calculation process of the MKP-ECA module is expressed as:
[0072]
[0073] where w k1 , w k2 , w k3 represent the convolutional weights with kernel sizes of 3, 5, and 7 respectively, σ is the SoftMax function; GAP represents the global pooling layer.
[0074] Figure 6 shows the flow chart of classifying and identifying small-sample interference signals using three serially deployed local-global fusion networks. Experiments show that the classification and identification effect is the best when three local-global fusion networks are connected in series. The recognition results are as Figure 7 shown. The recognition rates of 5 types of jamming signals are shown in the confusion matrix of the figure. Figure 7 In it, confusion matrix represents the confusion matrix; actual represents the actual value; predicted represents the predicted value. It can be seen from Figure 7 that the recognition rates of each interference signal are not exactly the same. Among them, the recognition rate of the tone interference signal is the highest, reaching 100%, and the recognition rates of the remaining interference signals are all above 95%.
Claims
1. A method for classifying and identifying small-sample interference signals based on a local-global fusion network, characterized in that Including the following steps: Step 1: Construct a dataset Collect interference signals, assign different labels according to their categories, extract the in-phase component I and quadrature component Q of the interference signals, construct a dataset by taking the modulus of the signals, then perform normalization preprocessing on the data in the constructed dataset to obtain an original interference signal dataset, randomly select a set percentage of the data from the original interference signal dataset as a small-sample interference signal dataset, and divide the small-sample interference signal dataset into a training set and a test set; Step 2: Construct a local-global fusion network model, and use the training set and test set to iteratively train the constructed local-global fusion network model to obtain a trained local-global fusion network model; the local-global fusion network model includes a multi-scale dilated convolution module, a multi-head self-attention module, and a multi-convolution kernel efficient channel attention module; the local-global fusion network model is configured as follows: First, the multi-scale dilated convolution module and the multi-head self-attention module are arranged in parallel to extract local features and global features respectively; Subsequently, the output features of the multi-scale dilated convolution module and the multi-head self-attention module are stacked on the channel dimension respectively to obtain an output feature map with local and global information. Where and are the output features of the multi-scale dilated convolution module and the multi-head self-attention module respectively, C represents the number of channels of the feature map, and T represents the time dimension of the feature map; Finally, the output feature map with local and global information is input into the multi-convolution kernel efficient channel attention module. Let the input feature vector be In the multi-convolution kernel efficient channel attention module, first, a global pooling layer is used to compress the input features into aggregated channels to obtain a representative value for each channel; Secondly, three convolution weights with different kernel sizes are used to scale and weight the aggregated channels; Activated by the Sigmoid function to obtain a scaling channel weight in the range of 0-1. The scaling channel weight is multiplied element-wise with the input feature vector to obtain the output feature vector The calculation process of the multi-convolution kernel efficient channel attention module is expressed as: where w k1 , w k2 and w k3 represent the convolution weights with kernel sizes of 3, 5, and 7 respectively, σ is the SoftMax function; GAP represents the global pooling layer; Step 3: Input the small-sample interference signal to be recognized into the trained local-global fusion network model for classification and recognition to obtain a recognition result.
2. The few-shot interference signal classification and recognition method based on the local-global fusion network according to claim 1, wherein Take 8% of the sample quantity of the original interference signal dataset as the small-sample interference signal dataset.
3. The small-sample interference signal classification and recognition method based on the local-global fusion network according to claim 1, wherein The multi-scale dilated convolution module is configured as follows: Adopt a three-parallel-branch structure; subsequently, stack the output elements of the three parallel branches on the channel using the Concat function to form an overall feature map; and use a convolutional layer with a kernel of 3 to solve the aliasing problem caused by the concatenation operator; this process is expressed as follows: Among them represents the input feature vector; represents the output feature of the multi-scale dilated convolution module, C represents the number of channels of the feature map, and T represents the input time dimension; represents the weight of the dilated convolution with a dilation rate of r, and its convolution kernel size is 5; represents the convolution weight with a kernel of 3, H r represents the output features of convolutions with different dilation rates.
4. The few-shot interference signal classification and recognition method based on the local-global fusion network according to claim 3, wherein The dilation rates adopted by the three parallel branches are 1, 2, and 5 respectively.
5. The small-sample interference signal classification and recognition method based on the local-global fusion network according to claim 1, characterized in that The multi-head self-attention module is configured as follows: Let the input feature vector be where C represents the number of channels of the feature map and T represents the time dimension of the input; First, perform feature mapping through three parallel linear transformations to obtain the query matrix φ h , the key matrix τ h and the value matrix ν h , This process is shown as follows: Among them, and respectively represent the linear mapping weights of the query matrix, key matrix, and value matrix; C h represents the feature dimension, and the query matrix, key matrix, and value matrix have the same feature dimension; h represents the number of heads of the multi-head self-attention module, where C = h × C h ; Subsequently, perform a dot product operation on the query matrix φ h and the transpose of the key matrix τ h to obtain the similarity value between the feature maps; then perform scaling and normalization operations on it to obtain the attention weights, and further multiply these attention weights with the value matrix ν h to obtain the weighted features The process is shown as follows: Where σ is the SoftMax function, representing the normalization operation, and tr represents the matrix transpose; Stack the weighted features output by all the heads on the channels and perform feature mapping to obtain the output features It is expressed as follows: Among them represents the output feature of the multi-head self-attention module; represents the convolution weight with a kernel of 3.