Classification method and system for motor imagery electroencephalogram signals, medium and equipment
By using multi-scale temporal feature extraction module, spatial self-attention module and classification module of the Kolmogolov-Arnold network in the decoding model of motion imagination EEG signal, the problems of large amount of parameters and high computational complexity are solved, and more efficient and lightweight decoding performance is achieved.
Patent Information
- Application Number
- CN202510231473.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-28
- Publication Date
- 2025-05-16
AI Technical Summary
The existing CNN-based motion imaginary EEG signal decoding model has problems with large amounts of parameters and high training time complexity, which leads to difficulty in deploying on embedded devices with tight memory and computing resources, limiting the portability and integration of the model.
The multi-scale time feature extraction module, spatial self-attention module and the classification module of the Kolmogolov-Arnold network are used to extract features through depth separation convolution and spatial convolution layers, and replace the fully connected layer with KAN to reduce the number of layers and parameter amount of the model.
On the premise of ensuring decoding performance, the complexity and computing complexity of the model are reduced, the decoding efficiency and the lightweighting of the model are improved, making it more suitable for deployment on low-computing equipment.
Smart Images
Figure CN120011892A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of electroencephalogram (EEG) signal classification, and in particular to a method, system, medium and device for classifying motor imagery EEG signals. Background Art
[0002] Motor imagery brain-computer interface technology can be widely used in various fields. The premise for creating more valuable brain control systems is to ensure that the motor imagery electroencephalogram (MI-EEG) signal has high recognition accuracy. Initially, traditional machine learning methods were widely used to classify MI-EEG signals, including: common spatial patterns, universal sparse spectral spatial patterns, filter bank common spatial patterns, etc.
[0003] Although the performance of traditional motor imagery EEG decoding has been greatly improved, the quality of feature extraction of these methods depends largely on human prior knowledge, which limits the performance of traditional EEG classification methods. In recent years, deep learning methods have made breakthrough progress in the field of motor imagery EEG decoding and have been widely used. Convolutional Neural Networks (CNN) can automatically learn general features, specific features, and potential complex features in EEG signals.
[0004] However, at present, the CNN-based MI-EEG signal decoding model has an increasing number of convolutional layers and input neurons in the fully connected layer, which has the problems of large number of parameters and high training time complexity, making it difficult to deploy on embedded devices with limited memory and computing resources. This limits the portability and integration of the MI-EEG signal decoding model. Summary of the invention
[0005] The purpose of the present invention is to overcome the deficiencies in the prior art and to provide a method, system, medium and device for classifying motor imagery EEG signals, which can minimize the number of layers and parameters of the model to the greatest extent while ensuring decoding performance, so as to reduce the complexity of the model and improve decoding efficiency.
[0006] To achieve the above object, the present invention is implemented by adopting the following technical solutions: On the one hand, the present invention provides a method for classifying motor imagery EEG signals, comprising: Obtaining raw EEG signals; Inputting the original EEG signal into a pre-built EEG signal classification model, and outputting the classification result of the EEG signal; Wherein, the EEG signal classification model includes a multi-scale temporal feature extraction module, a spatial self-attention module and a classification module connected in sequence; The multi-scale temporal feature extraction module includes a plurality of parallel depth-separable convolutional layers, and the convolution kernel size of each depth-classifiable convolutional layer is different; The spatial self-attention module includes a plurality of parallel spatial convolutional layers, and the number of the spatial convolutional layers is the same as the number of the depth-separable convolutional layers; In the classification module, the fully connected layers are replaced by Kolmogorov-Arnold networks; After the depthwise separable convolutional layer is connected to the spatial convolutional layer in a one-to-one correspondence, it is connected to the classification module.
[0007] Optionally, before inputting the original EEG signal into a pre-built EEG signal classification model, the method further includes: The original EEG signal is sequentially subjected to downsampling processing, filtering processing and matrix format conversion to obtain an EEG signal in matrix format.
[0008] Optionally, the processing steps of the EEG signal classification model include: In the depthwise separable convolutional layer, the time dimension features of the original EEG signal are extracted to obtain a time feature map; In the spatial convolution layer, the spatial dimension features of the temporal feature map are extracted to obtain a spatiotemporal feature map; In the classification module, the spatiotemporal feature graph is sequentially spliced, flattened, and classified to obtain a classification result of the EEG signal; wherein the Kolmogorov-Arnold network is used to classify the spatiotemporal feature graph.
[0009] Optionally, extracting the time dimension feature of the original EEG signal to obtain a time feature graph includes: (1); (2); in, Indicated in On the channel Channel-wise convolution value of position output; Indicated in On the channel Temporal feature map of position output; Indicates the size of the convolution kernel; Indicates input channel The original EEG signal on Indicates channel The convolution kernel on ; represents the point-wise convolution kernel; Indicates the number of input channels for channel-by-channel convolution; Indicates the number of output channels for channel-by-channel convolution and the number of input channels for point-by-point convolution; Indicates the number of output channels of point-by-point convolution.
[0010] Optionally, before extracting the spatial dimension features of the time feature graph, the method further includes: (3); in, Represents the normalized time feature map; is the time characteristic diagram; is the mean of the time feature map; is the variance of the temporal feature map; is a constant.
[0011] Optionally, extracting the spatial dimension features of the temporal feature map to obtain the spatiotemporal feature map includes: (4); (5); (6); in, represents the global pooling vector; Indicates channel The weight of Represents a spatiotemporal feature map; Represents spatial dimension; Indicates that in the channel Position on Output spatiotemporal feature map; Represents the convolution kernel size of one-dimensional convolution; represents one-dimensional convolution; express sigmoid Activation function; Represents the number of spatiotemporal channels after attention weighting.
[0012] Optionally, classifying and identifying the output feature map to obtain a classification result of the EEG signal includes: (7); (8); in, represents an intermediate variable; Indicates the classification result of EEG signal; Represents the one-dimensional feature number of the flattened feature map; represents the weight of the linear combination; Represents the flattened feature map; represents the bias term; represents the number of neurons in the hidden layer of the Kolmogorov-Arnold network; represents the linear weighting coefficient; represents a non-linear activation function.
[0013] In a second aspect, the present invention provides a classification system for motor imagery EEG signals, comprising: Signal acquisition module, to obtain the original EEG signal; A signal classification module, inputting the original EEG signal into a pre-built EEG signal classification model, and outputting the classification result of the EEG signal; Wherein, the EEG signal classification model includes a multi-scale temporal feature extraction module, a spatial self-attention module and a classification module connected in sequence; The multi-scale temporal feature extraction module includes a plurality of parallel depth-separable convolutional layers, and the convolution kernel size of each depth-classifiable convolutional layer is different; The spatial self-attention module includes a plurality of parallel spatial convolutional layers, and the number of the spatial convolutional layers is the same as the number of the depth-separable convolutional layers; In the classification module, the fully connected layers are replaced by Kolmogorov-Arnold networks; After the depthwise separable convolutional layer is connected to the spatial convolutional layer in a one-to-one correspondence, it is connected to the classification module.
[0014] In a third aspect, the present invention provides a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the steps of the method described in the first aspect.
[0015] In a fourth aspect, the present invention provides a computer device, including a processor and a storage medium; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the method according to the first aspect.
[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. Under the premise of ensuring decoding performance, the network model of the present invention makes full use of the low parameter amount and low computational complexity of the depthwise separable convolution and Kolmogorov-Arnold network, achieves the advantages of lightweight and efficient computation, and minimizes the number of layers and parameters of the model to reduce the complexity of the model and improve decoding efficiency; 2. In terms of model storage, the existing EEG signal deep learning decoding methods have the disadvantages of complex structure and low reasoning efficiency. The present invention is based on a shallow CNN built on the basis of deep separable convolution and Kolmogorov-Arnold network. The number of parameters and layers of the model are greatly reduced, which significantly reduces the storage space required for the model, making it more lightweight and more suitable for scenarios requiring efficient storage. 3. In terms of computational efficiency, most existing network models are constructed by combining the self-attention mechanism and the fully connected layer, but this method has a high computational complexity. The present invention adopts the channel attention + spatial convolution method to process the spatial and channel information separately compared to the ordinary convolution method, which can greatly reduce the amount of calculation and parameters, improve the decoding efficiency, and significantly improve the spatial feature expression ability of the model while maintaining a low computational overhead. In addition, the classification method of using KAN instead of the fully connected layer reduces the large-scale matrix multiplication operations of the fully connected layer, improves the speed and computational efficiency of network reasoning, and makes the model easier to deploy on low-computing power devices. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] Figure 1 FIG. 1 is a schematic diagram showing a structure of a method for classifying motor imagery EEG signals in one embodiment of the present invention; Figure 2 Shown is a schematic structural diagram of another embodiment of the method for classifying motor imagery EEG signals of the present invention. DETAILED DESCRIPTION
[0018] The technical solution of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations on the technical solution of the present invention. The embodiments of the present invention and the technical features in the embodiments may be combined with each other unless there is a conflict.
[0019] The term "and / or" is only a description of the association relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " generally indicates that the related objects are in an "or" relationship.
[0020] Example 1
[0021] like Figure 1 As shown, this embodiment introduces a classification method for motor imagery EEG signals, builds a lightweight and efficient EEG signal classification model, and minimizes the number of layers and parameters of the model to the greatest extent while ensuring decoding performance, so as to reduce the complexity of the model and improve decoding efficiency. Specifically, the following steps are included: Step 1: Get the original EEG signal and perform simple preprocessing operations on the original EEG signal, specifically: First, the original EEG signal is downsampled to 250 Hz to reduce the amount of calculation and improve the efficiency of the algorithm.
[0022] Next, the signal is filtered, including a bandpass filter from 0.5 Hz to 100 Hz and a notch filter at 50 Hz.
[0023] Finally, convert the signal to a matrix format of (1, channel, time point).
[0024] Step 2: Input the original EEG signal into a pre-built EEG signal classification model, and output the classification result of the EEG signal; wherein, Figure 2 As shown, the EEG signal classification model includes a multi-scale time feature extraction module, a spatial self-attention module and a classification module in sequence; The multi-scale temporal feature extraction module is responsible for temporal filtering of the original EEG signals and extraction of global and local features; the spatial self-attention module is responsible for the integration of spatial information such as channel importance and interaction between channels; the improved classification module is responsible for converting the extracted temporal and spatial features into specific categories.
[0025] In a specific embodiment, the constructed EEG signal classification model is a Multi-scale Spatio-Temporal Attention Convolutional Neural Network (MSTACNN), and the Kolmogorov-Arnold Network (KAN) is introduced in the classification module to replace the fully connected layer for recognition and classification. Figure 2 shown.
[0026] like Figure 1 As shown, the processing steps of the EEG signal classification model include: In the multi-scale time feature extraction module, the multi-scale time feature extraction module includes multiple parallel depth separable convolutional layers, and the convolution kernel size of each depth classifiable convolutional layer is different. In this embodiment, three depth separable convolutional layers with different kernel sizes are constructed to perform preliminary time dimension feature extraction on the original EEG signal to obtain time feature maps with different receptive field sizes, specifically: In this embodiment, in order to enable the network model to learn the global and detailed temporal features, three depth-wise separable convolutional layers with different kernel sizes are constructed. The kernel sizes are (1, 128), (1, 32), and (1, 8), respectively. The number of filters is 8, the padding is 0, and the step size is 1. It can be adjusted according to the specific task requirements, by selecting different convolution kernel sizes, and increasing or decreasing the number of branches of the depth-wise separable convolutional layer to replace the original construction method, so as to better adapt to the feature requirements of the task.
[0027] Depthwise separable convolution is a convolution operation that improves the computational efficiency of convolutional neural networks. First, convolution is performed channel by channel: a convolution kernel is applied independently to each input channel, as shown in the calculation formula (1): (1); Then, point-by-point convolution: use a 1×1 convolution kernel to perform a linear combination of channels at each pixel, as shown in calculation formula (2): (2); in, Indicated in On the channel Channel-wise convolution value of position output; Indicated in On the channel Temporal feature map of position output; Indicates the size of the convolution kernel; Indicates input channel The original EEG signal on Indicates channel The convolution kernel on ; Represents the point-by-point convolution kernel; Indicates the number of input channels for channel-by-channel convolution; Indicates the number of output channels for channel-by-channel convolution and the number of input channels for point-by-point convolution; Indicates the number of output channels of point-by-point convolution.
[0028] Compared with the method of processing spatial and channel information together by ordinary convolution, this method of processing spatial and channel information separately can greatly reduce the amount of calculation and parameters and improve decoding efficiency.
[0029] Before performing spatial convolution, batch normalization and channel attention operations are performed. Specifically, batch normalization uses training batches as units to normalize data to a normal distribution with a mean of 0 and a variance of 1, so that each layer of the network learns the same distribution, accelerates network convergence, and improves network generalization ability, as shown in formula (3): (3); in, Represents the normalized time feature map; is the time characteristic diagram; is the mean of the time feature map; is the variance of the temporal feature map; is a constant.
[0030] In the spatial self-attention module, the temporal feature map is subjected to spatial dimension feature extraction to obtain a spatiotemporal feature map that integrates spatiotemporal information, specifically: The spatial self-attention module includes a plurality of parallel spatial convolutional layers, the number of which is the same as that of the depthwise separable convolutional layers; the depthwise separable convolutional layers are connected to the spatial convolutional layers in a one-to-one correspondence; the spatial convolutional layers can allow the model to better focus on specific channels by assigning different weights to different channels, thereby improving the performance of the model in specific tasks. First, global average pooling is performed: the spatial information of each channel is compressed into a scalar to obtain a vector containing the global information of all channels, as shown in calculation formula (4): (4); Then, 1D convolution: directly perform convolution operation on the vector after global pooling to capture the local interaction relationship between channels, as shown in calculation formula (5): (5); Finally, reweight the channels: apply the attention weights to each input channel and perform a channel-by-channel weighting operation, as shown in the calculation formula (6): (6); in, represents the global pooling vector; Indicates channel The weight of Represents a spatiotemporal feature map; Represents spatial dimension; Indicates that in the channel Position on Output spatiotemporal feature map; Represents the convolution kernel size of one-dimensional convolution; represents one-dimensional convolution; express sigmoid Activation function; Represents the number of spatiotemporal channels after attention weighting.
[0031] The spatial convolution layer can capture the spatial dependencies of EEG signals in the channel dimension, fuse the features between different channels, reduce the dimension, reduce the amount of calculation, and automatically learn effective neural spatial features, thereby improving the model's perception and feature expression capabilities of activities in different brain regions.
[0032] Different branches have the same spatial convolution operation, the convolution kernel size is (C, 1), the number of filters is 16, and the padding is 0.
[0033] In the classification module, the three spatiotemporal feature maps are first concatenated, and then batch normalization, activation function ELU, average pooling, Dropout operation, and flattening are performed in sequence to obtain the flattened feature map. , and then input the flattened feature map into the KAN layer, and finally identify the classification result of the EEG signal, which is: Batch normalization is the same as formula (3). The activation function ELU retains the advantages of ReLU. At the same time, through negative value processing and exponential decay, it solves the limitations of ReLU in the negative range and enhances the learning ability and robustness of the neural network. The kernel size of the average pooling is (1, 4), the step size is 4, and the Dropout rate is set to 0.25.
[0034] KAN is a new neural network architecture based on the Kolmogorov-Arnold representation theorem, which aims to solve the representation problem of high-dimensional functions. KAN uses function decomposition theory to decompose multidimensional problems into a series of single-variable functions and simple linear combinations, thereby effectively reducing computational complexity and the number of network parameters.
[0035] Compared with the traditional fully connected layer, KAN has unique advantages in processing high-dimensional data. The main idea of KAN is to use function decomposition to replace the matrix operation of the traditional fully connected layer. First, the input mapping layer: a set of linear combinations are mapped into several intermediate variables, such as formula (7): (7); Then, the nonlinear activation layer: each intermediate variable Through non-linear activation function To process: , and finally all output Through linear weighted combination, the classification results of EEG signals are obtained , as shown in formula (8): (8); in, Represents the one-dimensional feature number of the flattened feature map; represents the weight of the linear combination; Represents the flattened feature map; represents the bias term; represents the number of neurons in the hidden layer of the Kolmogorov-Arnold network; Represents the linear weighting coefficient.
[0036] KAN approximates high-dimensional functions , without the need for a large number of parameters and calculations in traditional fully connected networks.
[0037] Finally, the probability of each category is calculated through the Softmax activation function, and the category with the highest probability is selected as the final prediction result.
[0038] This embodiment proposes a classification method for motor imagery EEG signals. Under the premise of ensuring decoding performance, the network model fully utilizes the low parameter amount and low computational complexity of deep separable convolution and KAN, and achieves the advantages of lightweight and efficient computation.
[0039] Example 2
[0040] Based on the same inventive concept as Example 1, this example introduces a classification system for motor imagery EEG signals, including: Signal acquisition module, to obtain the original EEG signal; A signal classification module, inputting the original EEG signal into a pre-built EEG signal classification model, and outputting the classification result of the EEG signal; Wherein, the EEG signal classification model includes a multi-scale temporal feature extraction module, a spatial self-attention module and a classification module connected in sequence; The multi-scale temporal feature extraction module includes a plurality of parallel depth-separable convolutional layers, and the convolution kernel size of each depth-classifiable convolutional layer is different; The spatial self-attention module includes a plurality of parallel spatial convolutional layers, and the number of the spatial convolutional layers is the same as the number of the depth-separable convolutional layers; In the classification module, the fully connected layers are replaced by Kolmogorov-Arnold networks; After the depthwise separable convolutional layer is connected to the spatial convolutional layer in a one-to-one correspondence, it is connected to the classification module.
[0041] The specific functional implementation of each of the above modules can be found in the relevant contents of the method in Example 1 and will not be elaborated here.
[0042] Example 3
[0043] This embodiment introduces a computer-readable storage medium on which computer instructions are stored. When the computer instructions are executed by a processor, the steps of the method described in Embodiment 1 are implemented.
[0044] Example 4
[0045] This embodiment introduces a computer device, including a processor and a storage medium; The storage medium is used to store instructions; The processor is used to operate according to the instruction to execute the method according to embodiment 1.
[0046] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.
[0047] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0048] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0049] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.
[0050] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the enlightenment of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the purpose of the present invention and the claims, which all fall within the protection of the present invention.
Claims
1. A method for classifying motor imagery EEG signals, characterized in that: include: Obtaining raw EEG signals; Inputting the original EEG signal into a pre-built EEG signal classification model, and outputting the classification result of the EEG signal; Wherein, the EEG signal classification model includes a multi-scale temporal feature extraction module, a spatial self-attention module and a classification module connected in sequence; The multi-scale temporal feature extraction module includes a plurality of parallel depth-separable convolutional layers, and the convolution kernel size of each depth-classifiable convolutional layer is different; The spatial self-attention module includes a plurality of parallel spatial convolutional layers, and the number of the spatial convolutional layers is the same as the number of the depth-separable convolutional layers; In the classification module, the fully connected layers are replaced by Kolmogorov-Arnold networks; After the depthwise separable convolutional layer is connected to the spatial convolutional layer in a one-to-one correspondence, it is connected to the classification module.
2. The method for classifying motor imagery EEG signals according to claim 1, characterized in that: Before inputting the original EEG signal into the pre-built EEG signal classification model, the method further includes: The original EEG signal is sequentially subjected to downsampling processing, filtering processing and matrix format conversion to obtain an EEG signal in matrix format.
3. The method for classifying motor imagery EEG signals according to claim 1, characterized in that: The processing steps of the EEG signal classification model include: In the depthwise separable convolutional layer, the time dimension features of the original EEG signal are extracted to obtain a time feature map; In the spatial convolution layer, the spatial dimension features of the temporal feature map are extracted to obtain a spatiotemporal feature map; In the classification module, the spatiotemporal feature graph is sequentially spliced, flattened, and classified to obtain a classification result of the EEG signal; wherein the Kolmogorov-Arnold network is used to classify the spatiotemporal feature graph.
4. The method for classifying motor imagery EEG signals according to claim 3, characterized in that: Extracting the time dimension features of the original EEG signal to obtain a time feature graph includes: （1）； （2）； in, Indicated in On the channel Channel-wise convolution value of position output; Indicated in On the channel Temporal feature map of position output; Indicates the size of the convolution kernel; Indicates input channel The original EEG signal on Indicates channel The convolution kernel on ; represents the point-wise convolution kernel; Indicates the number of input channels for channel-by-channel convolution; Indicates the number of output channels for channel-by-channel convolution and the number of input channels for point-by-point convolution; Indicates the number of output channels of point-by-point convolution.
5. The method for classifying motor imagery EEG signals according to claim 3, characterized in that: Before extracting the spatial dimension features of the temporal feature graph, the method further includes: （3）； in, Represents the normalized time feature map; is the time characteristic diagram; is the mean of the time feature map; is the variance of the temporal feature map; is a constant.
6. The method for classifying motor imagery EEG signals according to claim 5, characterized in that: Extracting the spatial dimension features of the temporal feature graph to obtain a spatiotemporal feature graph includes: （4）； （5）； （6）； in, represents the global pooling vector; Indicates channel The weight of Represents a spatiotemporal feature map; Represents spatial dimension; Indicates that in the channel Position on Output spatiotemporal feature map; Represents the convolution kernel size of one-dimensional convolution; represents one-dimensional convolution; express sigmoid Activation function; Represents the number of spatiotemporal channels after attention weighting.
7. The method for classifying motor imagery EEG signals according to claim 3, characterized in that: Classifying and identifying the output feature map to obtain a classification result of the EEG signal includes: （7）； （8）； in, represents an intermediate variable; Indicates the classification result of EEG signal; Represents the one-dimensional feature number of the flattened feature map; represents the weight of the linear combination; Represents the flattened feature map; represents the bias term; represents the number of neurons in the hidden layer of the Kolmogorov-Arnold network; represents the linear weighting coefficient; represents a non-linear activation function.
8. A classification system for motor imagery EEG signals, characterized in that: include: Signal acquisition module, to obtain the original EEG signal; A signal classification module, inputting the original EEG signal into a pre-built EEG signal classification model, and outputting the classification result of the EEG signal; Wherein, the EEG signal classification model includes a multi-scale temporal feature extraction module, a spatial self-attention module and a classification module connected in sequence; The multi-scale temporal feature extraction module includes a plurality of parallel depth-separable convolutional layers, and the convolution kernel size of each depth-classifiable convolutional layer is different; The spatial self-attention module includes a plurality of parallel spatial convolutional layers, and the number of the spatial convolutional layers is the same as the number of the depth-separable convolutional layers; In the classification module, the fully connected layers are replaced by Kolmogorov-Arnold networks; After the depthwise separable convolutional layer is connected to the spatial convolutional layer in a one-to-one correspondence, it is connected to the classification module.
9. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the computer instructions are executed by a processor, the steps of the method described in any one of claims 1 to 7 are implemented.
10. A computer device, characterized in that: including processor and storage medium; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the method according to any one of claims 1 to 7.