CN116058852B - Classification system, method, electronic device and storage medium for MI-EEG signals - Google Patents

Abstract

The embodiment of the invention discloses a classification system, a classification method, electronic equipment and a storage medium of MI-EEG signals, wherein the classification method of the MI-EEG signals comprises the following steps: preprocessing an MI-EEG signal data set to obtain a preprocessed data set, and dividing a training set and a testing set based on the data set; constructing a multi-branch convolutional neural network model, wherein the first branch, the second branch and the third branch respectively comprise an EEGNet convolutional block and a convolutional attention module; obtaining a target multi-branch convolutional neural network model through a training set and a testing set; and inputting the MI-EEG signals to be classified into a target multi-branch convolutional neural network model to obtain a classification result. The MI-EEG signal classification method solves the problem that the existing traditional model can not finish specific classification tasks while having fixed super parameters.

Description

MI-EEG signal classification system, method, electronic device and storage medium

Technical field

The present invention relates to the technical field of MI-EEG signal classification, and specifically relates to a MI-EEG signal classification system, method, electronic equipment and storage medium.

Background technique

Motor imagery (MI) is a behavior of imagining body movement, which is generated when the brain imagines moving a specific limb. It can be captured and identified through electroencephalography (EEG); in the early stage, common spatial patterns (common spatial patterns) spatial patterns (CSP) is the best technology to identify EEG-MI based signals.

The decoding of EEG-MI signals has potential application value in fields such as sports medicine. However, due to the low signal-to-noise ratio of EEG-MI signals, the presence of motion artifacts and noise, and the spatial correlation of signals, classification becomes very challenging. characteristics, and the EGG-MI signal has extremely high specificity, with great differences between different subjects, and may also have great differences at different test time points of the same subject. Such a specialized task is not suitable for traditional models. It consumes a lot of time and computational cost.

Contents of the invention

The purpose of the embodiments of the present invention is to provide a MI-EEG signal classification system, method, electronic device and storage medium to solve the problem that the existing traditional model cannot complete specific classification tasks while having fixed hyperparameters.

To achieve the above objectives, embodiments of the present invention provide a method for classifying MI-EEG signals. The method specifically includes:

Obtain a MI-EEG signal data set, perform preprocessing on the MI-EEG signal data set to obtain a preprocessed data set, and divide a training set and a test set based on the data set;

Construct a multi-branch convolutional neural network model, wherein the multi-branch convolutional neural network model includes a first branch, a second branch and a third branch, the first branch, the second branch and the third branch They include EEGNet convolution blocks and convolution attention modules respectively, and the first branch, the second branch and the third branch are connected through a softmax layer;

Train the multi-branch convolutional neural network model through the training set;

Perform performance evaluation on the trained multi-branch convolutional neural network through the test set to obtain the target multi-branch convolutional neural network model;

The MI-EEG signal to be classified is input into the target multi-branch convolutional neural network model to obtain a classification result.

On the basis of the above technical solutions, the present invention can also make the following improvements:

Further, the EEGNet convolution block includes a first convolution layer, a second convolution layer and a third convolution layer connected in sequence, and the first convolution layer, the second convolution layer and the The window size of the third convolutional layer is different.

Further, the convolution attention module includes a channel attention sub-module and a spatial attention sub-module;

The channel attention sub-module includes an average pooling layer, a maximum pooling layer and a shared network. The shared network includes a multi-layer perceptron, the multi-layer perceptron has a hidden layer, the average pooling layer and the maximum pooling layer. The layer generates a feature map after receiving input features, and the sharing network outputs channel-refined features after receiving the feature map;

The spatial attention sub-module includes an average pooling layer, a maximum pooling layer and a convolutional layer, and the channel-refined features are refined through the output of the average pooling layer, the maximum pooling layer and the convolutional layer. Feature mapping.

Further, the construction of a multi-branch convolutional neural network model, wherein the multi-branch convolutional neural network model includes a first branch, a second branch and a third branch, the first branch, the second branch and The third branch includes an EEGNet convolution block and a convolution attention module respectively. The first branch, the second branch and the third branch are connected through a softmax layer, including:

The EEGNet convolution block and convolution attention module in the first branch, the second branch and the third branch are respectively set with different numbers of parameters to capture different features.

A classification system for MI-EEG signals, including:

Acquisition module, used to obtain MI-EEG signal data sets;

A preprocessing module, used to preprocess the MI-EEG signal data set to obtain a preprocessed data set, and divide a training set and a test set based on the data set;

A building module for building a multi-branch convolutional neural network model, wherein the multi-branch convolutional neural network model includes a first branch, a second branch and a third branch, the first branch, the second branch and The third branch includes an EEGNet convolution block and a convolution attention module respectively, and the first branch, the second branch and the third branch are connected through a softmax layer;

A training module, used to train the multi-branch convolutional neural network model through the training set;

A test module, used to evaluate the performance of the trained multi-branch convolutional neural network through the test set to obtain the target multi-branch convolutional neural network model;

The multi-branch convolutional neural network model obtains classification results based on the MI-EEG signal to be classified.

Further, the EEGNet convolution block includes a first convolution layer, a second convolution layer and a third convolution layer connected in sequence, and the first convolution layer, the second convolution layer and the The window size of the third convolutional layer is different.

Further, the convolution attention module includes a channel attention sub-module and a spatial attention sub-module;

The channel attention sub-module includes an average pooling layer, a maximum pooling layer and a shared network. The shared network includes a multi-layer perceptron, the multi-layer perceptron has a hidden layer, the average pooling layer and the maximum pooling layer. The layer generates a feature map after receiving input features, and the sharing network outputs channel-refined features after receiving the feature map;

The spatial attention sub-module includes an average pooling layer, a maximum pooling layer and a convolutional layer, and the channel-refined features are refined through the output of the average pooling layer, the maximum pooling layer and the convolutional layer. Feature mapping.

Further, the MI-EEG signal classification system further includes a setting module for converting the EEGNet convolution blocks and convolution blocks in the first branch, the second branch and the third branch. The attention module sets different numbers of parameters respectively to capture different features.

An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the method are implemented.

A non-transitory computer-readable storage medium on which a computer program is stored, which implements the steps of the method when executed by a processor.

The embodiments of the present invention have the following advantages:

The MI-EEG signal classification method in the present invention obtains a MI-EEG signal data set, preprocesses the MI-EEG signal data set to obtain a preprocessed data set, and divides a training set and a test set based on the data set. ;Construct a multi-branch convolutional neural network model, wherein the multi-branch convolutional neural network model includes a first branch, a second branch and a third branch, the first branch, the second branch and the third branch The branches respectively include EEGNet convolution blocks and convolution attention modules, and the first branch, the second branch and the third branch are connected through a softmax layer; the multi-branch convolutional neural network is tested through the training set. The model is trained; the performance of the trained multi-branch convolutional neural network is evaluated through the test set to obtain a target multi-branch convolutional neural network model; the MI-EEG signal to be classified is input into the target multi-branch convolutional neural network network model to obtain classification results; it solves the problem that existing traditional models cannot complete specific classification tasks while having fixed hyperparameters.

Description of the drawings

In order to more clearly explain the embodiments of the present invention or the technical solutions in the prior art, the drawings that need to be used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are only exemplary. For those of ordinary skill in the art, other implementation drawings can be obtained based on the extension of the provided drawings without exerting creative efforts.

The structures, proportions, sizes, etc. shown in this specification are only used to coordinate with the contents disclosed in the specification for the understanding and reading of people familiar with this technology. They are not used to limit the conditions under which the invention can be implemented, and therefore do not have any technical Any structural modification, change in proportion or size adjustment shall still fall within the scope of the technical content disclosed in the present invention without affecting the effectiveness and purpose achieved by the present invention. within the scope that can be covered.

Figure 1 is a flow chart of the MI-EEG signal classification method of the present invention;

Figure 2 is a first architecture diagram of the MI-EEG signal classification system of the present invention;

Figure 3 is a second architecture diagram of the MI-EEG signal classification system of the present invention;

Figure 4 is an architectural diagram of the EEGNet convolution block of the present invention;

Figure 5 is an architectural diagram of the convolutional attention module of the present invention;

Figure 6 is an architectural diagram of the channel attention sub-module of the present invention;

Figure 7 is an architectural diagram of the spatial attention sub-module of the present invention;

Figure 8 is an architectural diagram of the multi-branch convolutional neural network model of the present invention;

Figure 9 is a schematic diagram of the physical structure of the electronic equipment provided by the present invention.

The drawings are marked as:

Acquisition module 10, preprocessing module 20, construction module 30, training module 40, test module 50, setting module 60, electronic device 70, processor 701, memory 702, bus 703.

Detailed ways

The following specific embodiments are used to illustrate the implementation of the present invention. Persons familiar with this technology can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only part of the embodiments of the present invention. , not all examples. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

Example

Figure 1 is a flow chart of an embodiment of a MI-EEG signal classification method of the present invention. As shown in Figure 1, a MI-EEG signal classification method provided by an embodiment of the present invention includes the following steps:

S101, obtain the MI-EEG signal data set, preprocess the MI-EEG signal data set to obtain the preprocessed data set, and divide the training set and the test set based on the data set;

Specifically, the training set used 22 EEG electrodes to collect 4.5 s data with a sampling frequency of 250 Hz (250×4.5=1125 samples). Each trial produces a data matrix of dimension (22×1125). MI is divided into four categories: left hand, right hand, foot and tongue.

The validation set is collected using 128 channels with a sampling frequency of 500 Hz. The validation set was downsampled from 500 Hz to 250Hz to improve data quality. Reduce signal channels from 128 to 44, thereby excluding electrodes not connected to the MI region. In order to be consistent with the training set, each trial lasted 4.5 seconds (a prompt was given 0.5 seconds before the end of the trial). Each trial produced 1125 samples, and the data matrix was (44×1125).

S102. Construct a multi-branch convolutional neural network model. The multi-branch convolutional neural network model includes a first branch, a second branch and a third branch. The first branch, the second branch and the third branch respectively include EEGNet convolution blocks. And the convolutional attention module, the first branch, the second branch and the third branch are connected through the softmax layer;

Specifically, the EEGNet convolution block includes a first convolution layer, a second convolution layer and a third convolution layer connected in sequence, and the first convolution layer, the second convolution layer and the The third convolutional layer has a different window size, which is defined by the kernel size. The first convolutional layer uses 2D filters, followed by batch normalization. Batch normalization helps speed up training and regularize the model. The second convolutional layer uses depthwise convolutions with batch normalization and activation functions in the form of exponential linear units (ELU), average pooling, and hidden layers. The third convolutional layer uses separable convolutions. The simplified architecture of EEGNet is shown in Figure 4.

The Convolutional Block Attention Module (CBAM) is a module that can be added to the model to focus on specific attributes and ignore others to emphasize important features along channels and spatial axes. Each branch can learn features on the channel and spatial axes by using a sequence of attention modules (shown in Figure 5). Because the convolutional attention module learns what information to highlight or hide, the convolutional attention module can effectively help the flow of information in the network.

The convolutional attention module has two sub-modules: channel attention sub-module and spatial attention sub-module.

As shown in Figure 6, the channel attention sub-module includes an average pooling layer, a maximum pooling layer and a shared network. The shared network includes a multi-layer perceptron. The multi-layer perceptron has a hidden layer. The average pooling layer The layer and the maximum pooling layer generate a feature map after receiving input features, and the sharing network outputs channel-refined features after receiving the feature map;

As shown in Figure 7, the spatial attention sub-module includes an average pooling layer, a maximum pooling layer and a convolution layer, and the channel-refined features include an average pooling layer, the maximum pooling layer and the convolution layer. The cumulative layer outputs a refined feature map.

In the channel attention sub-module, the input features from the previous convolution block are simultaneously transferred to the average pooling layer and the max pooling layer. The feature maps generated by the two pooling layers are then transferred to the shared network, which consists of a multi-layer perceptron (MLP) with one hidden layer. In this hidden layer, a reduction ratio is used to reduce the number of activation maps and thus the parameters. After applying the shared network to each pooled feature map, the output feature maps are merged using element-wise summation. Then, to generate the input feature vector of the spatial attention sub-module, element-wise multiplication is used between the channel attention sub-module output feature map and the attention module input feature. When computing spatial attention feature maps, channel-axis average pooling and max pooling are used. Therefore, convolutional layers are used to construct effective feature descriptions.

Convolutional neural network (CNN) can solve the calculation problem of high-dimensional data (such as EEG signals) by performing convolution operations in a neural network environment. The convolution window is a small portion of the input neuron to which each neuron in the first hidden layer of the CNN is connected. All neurons have a bias and each connection has a weight. The window then slides across the entire input sequence, and each neuron in the hidden layer learns to analyze its specific aspects. Kernel size is the size or length of the convolution window. Instead of learning new weights and biases for each hidden layer neuron, CNN only learns a set of weights and a single bias applied to all hidden layer neurons, that is, weight sharing, which is calculated as follows:

in, is the hidden layer/> filter's/> The activation or output of a neuron,/> Corresponds to the activation function, /> is the shared overall bias of the filter, /> and/> is the core size,/> is a vector of shared weights,/> is the vector output by the preceding neuron, /> Represents the transpose operation.

In moving images, the optimal kernel size varies from object to object, and for the same object at different times. In order to solve the subject-specific problem in EEG-MI classification using CNN, the multi-branch convolutional neural network model has different kernel sizes for each branch and is able to find suitable convolution scales, that is, kernel sizes, for all objects. Using different kernel sizes helps the multi-branch convolutional neural network model complete subject-specific tasks and be more general.

The multi-branch convolutional neural network model has three branches that determine the convolution size, number of filters, hidden probabilities, and attention parameters for all data. At the same time, models can be customized to specific topics while increasing their applicability. In the first convolutional layer, based on local and global modulation, the model can learn temporal attributes and spatial attributes based on spatially distributed dissociation filters. For this purpose, the input data is represented as a 2D array, where the rows represent the number of electrodes and the columns represent the number of time steps.

The representation of the MI-EEG signal data set is:

in, is the number of trajectories,/> is the signal and its corresponding classification label,/> , of which/> is the number of categories. /> Represents the input signal (2D array),/> , of which/> Indicates the number of EEG channels,/> Indicates the length of EEG signal input.

The output of the classification system is the output from the last layer, a layer with a softmax activation function. The output of this layer is a vector containing the probability of each possible outcome or category. The sum of the probabilities of all possible outcomes or categories in the vector is 1. Define softmax as:

in is the input vector of the softmax function/> , which contains/> n elements of categories (results),/> is the />th in the input vector elements,/> and/> is the number of categories. The cost function or loss function is categorical cross entropy which takes the output probability from the softmax function and measures the distance from the true value which gives a value of 0 or 1 for each classification/result.

Using cross-entropy loss when adjusting model weights during training can minimize the loss. The smaller the loss, the better the model performance. The cross-entropy loss function is defined as:

for categories, among which/> is the true value,/> Is the first/> The softmax probability of each class,/> Calculated using base 2.

The multi-branch convolutional neural network model (Multi-Branch EGGNet model, MBEEGCBAM model) can be divided into the two parts explained above: EEGNet convolution block and CBAM module. The architecture of MBEEGCBAM is shown in Figure 8. It has three different branches. Each branch has an EEGNet convolution block, channel attention module and spatial attention module, and then connects them through a series layer. Each branch has a different number of parameters to capture different features.

S103, train the multi-branch convolutional neural network model through the training set;

Specifically, all training data in the training set use global parameters for model training. The parameters of the first branch are set as follows: the EEGNet convolution block uses the elu activation function, 4 sequential filters, the kernel size is 6, and the dropout rate is 0; the attention module uses the relu activation function, the ratio is 2, and the kernel size is 2. The parameters of the second branch are set as follows: the EEGNet convolution block uses the elu activation function, 8 temporal filters, the kernel size is 32, and the dropout rate is 0.1; the attention module uses the relu activation function, the ratio is 8, and the kernel size is 4. The parameters of the third branch are set as follows: the EEGNet convolution block uses the elu activation function, 16 sequential filters, the kernel size is 64, and the dropout rate is 0.2; the attention module uses the relu activation function, the ratio is 8, and the kernel size is 2.

During the training phase, use a callback at the end of each training epoch to save the best model weights based on the current best accuracy, and load the saved best model during the testing phase. The convolutional learning rate is 0.0009, the batch size is 64, and the number of training epochs is 1000. Using the Adam optimizer, the cost function is the cross-entropy error function.

S104. Use the test set to evaluate the performance of the trained multi-branch convolutional neural network to obtain the target multi-branch convolutional neural network model.

Specifically, accuracy is used to evaluate model performance. The MBEEGCBAM model has good average classification accuracy in both the training set and the validation set, which are 82.85% and 95.45% respectively, both higher than other traditional models.

S105, input the MI-EEG signal to be classified into the target multi-branch convolutional neural network model to obtain the classification result;

The multi-branch convolutional neural network model of the MI-EEG signal classification method of the present invention connects the features of these branches before using the softmax layer to classify multiple branches of the basic model, so that more features can be captured, thereby The EEG-MI signal is classified more accurately, and it shows good performance in both the training set and the validation set. Its average classification accuracy is 82.85% and 95.45% respectively.

The multi-branch convolutional neural network model uses global parameters for all test objects and references the convolutional attention module, making the feature map more refined at the channel and spatial levels, with higher accuracy, and as a general model It can complete tasks on specific topics at the same time, saving computing costs.

Figures 2 and 3 are flow charts of an embodiment of the MI-EEG signal classification system of the present invention; as shown in Figures 2 and 3, an MI-EEG signal classification system provided by an embodiment of the present invention includes the following steps:

Acquisition module, used to obtain MI-EEG signal data sets;

A preprocessing module, used to preprocess the MI-EEG signal data set to obtain a preprocessed data set, and divide a training set and a test set based on the data set;

A building module for building a multi-branch convolutional neural network model, wherein the multi-branch convolutional neural network model includes a first branch, a second branch and a third branch, the first branch, the second branch and The third branch includes an EEGNet convolution block and a convolution attention module respectively, and the first branch, the second branch and the third branch are connected through a softmax layer;

A training module, used to train the multi-branch convolutional neural network model through the training set;

A test module, used to evaluate the performance of the trained multi-branch convolutional neural network through the test set to obtain the target multi-branch convolutional neural network model;

The multi-branch convolutional neural network model obtains classification results based on the MI-EEG signal to be classified.

The EEGNet convolution block includes a first convolution layer, a second convolution layer and a third convolution layer connected in sequence, and the first convolution layer, the second convolution layer and the third convolution layer Stacked windows have different sizes.

The convolution attention module includes a channel attention sub-module and a spatial attention sub-module;

The channel attention sub-module includes an average pooling layer, a maximum pooling layer and a shared network. The shared network includes a multi-layer perceptron, the multi-layer perceptron has a hidden layer, the average pooling layer and the maximum pooling layer. The layer generates a feature map after receiving input features, and the sharing network outputs channel-refined features after receiving the feature map;

The spatial attention sub-module includes an average pooling layer, a maximum pooling layer and a convolutional layer, and the channel-refined features are refined through the output of the average pooling layer, the maximum pooling layer and the convolutional layer. Feature mapping.

The MI-EEG signal classification system also includes a setting module for converting the EEGNet convolution block and the convolution attention module in the first branch, the second branch and the third branch. Set different numbers of parameters respectively to capture different features.

A MI-EEG signal classification system of the present invention obtains a MI-EEG signal data set through an acquisition module; preprocesses the MI-EEG signal data set through a preprocessing module to obtain a preprocessed data set. The data set is divided into a training set and a test set; a multi-branch convolutional neural network model is constructed through building modules, wherein the multi-branch convolutional neural network model includes a first branch, a second branch and a third branch, and the first branch The branch, the second branch and the third branch respectively include an EEGNet convolution block and a convolutional attention module, and the first branch, the second branch and the third branch are connected through a softmax layer; through the The training set trains the multi-branch convolutional neural network model; the performance of the trained multi-branch convolutional neural network is evaluated through the test set to obtain the target multi-branch convolutional neural network model; the multi-branch convolutional neural network model is The product neural network model obtains the classification result based on the MI-EEG signal to be classified. It solves the problem that existing traditional models cannot complete specific classification tasks while having fixed hyperparameters.

Figure 9 is a schematic diagram of the physical structure of an electronic device provided by an embodiment of the present invention. As shown in Figure 9, the electronic device 70 includes: a processor 701 (processor), a memory 702 (memory) and a bus 703;

Among them, the processor 701 and the memory 702 complete communication with each other through the bus 703;

The processor 701 is used to call the program instructions in the memory 702 to execute the methods provided by the above method embodiments. For example, it includes: obtaining the MI-EEG signal data set, preprocessing the MI-EEG signal data set to obtain the predetermined data. The processed data set is divided into a training set and a test set based on the data set; a multi-branch convolutional neural network model is constructed, wherein the multi-branch convolutional neural network model includes a first branch, a second branch and a third branch. , the first branch, the second branch and the third branch respectively include EEGNet convolution block and convolution attention module, the first branch, the second branch and the third branch pass softmax layer connection; training the multi-branch convolutional neural network model through the training set; performing performance evaluation on the trained multi-branch convolutional neural network through the test set to obtain the target multi-branch convolutional neural network model; The MI-EEG signal to be classified is input into the target multi-branch convolutional neural network model to obtain a classification result.

This embodiment provides a non-transitory computer-readable storage medium. The non-transitory computer-readable storage medium stores computer instructions. The computer instructions cause the computer to execute the methods provided by the above method embodiments, for example, including: acquiring MI-EEG signals. A data set, preprocessing the MI-EEG signal data set to obtain a preprocessed data set, dividing a training set and a test set based on the data set; constructing a multi-branch convolutional neural network model, wherein the multi-branch The convolutional neural network model includes a first branch, a second branch and a third branch. The first branch, the second branch and the third branch respectively include an EEGNet convolution block and a convolution attention module. The first branch, the second branch and the third branch are connected through a softmax layer; the multi-branch convolutional neural network model is trained through the training set; the trained multi-branch convolutional neural network model is trained through the test set. Perform performance evaluation on the convolutional neural network to obtain a target multi-branch convolutional neural network model; input the MI-EEG signal to be classified into the target multi-branch convolutional neural network model to obtain a classification result.

Those of ordinary skill in the art can understand that all or part of the steps to implement the above method embodiments can be completed by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, It includes the steps of the above method embodiment; and the aforementioned storage medium includes: ROM, RAM, magnetic disk or optical disk and other various storage media that can store program codes.

The device embodiments described above are only illustrative. The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they may be located in one place. , or it can be distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Persons of ordinary skill in the art can understand and implement the method without any creative effort.

Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the part of the above technical solution that essentially contributes to the existing technology can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM/RAM, magnetic disc, optical disk, etc., including a number of instructions to cause a computer device (which can be a personal computer, a server, or a network device, etc.) to execute various embodiments or methods of certain parts of the embodiments.

Although the present invention has been described in detail with general descriptions and specific examples above, it is obvious to those skilled in the art that some modifications or improvements can be made on the basis of the present invention. Therefore, these modifications or improvements made without departing from the spirit of the present invention all fall within the scope of protection claimed by the present invention.

Claims (8)

1. A method of classifying MI-EEG signals, the method comprising in particular:

acquiring a M I-EEG signal data set, preprocessing the M I-EEG signal data set to obtain a preprocessed data set, and dividing a training set and a testing set based on the preprocessed data set;

constructing a multi-branch convolutional neural network model, wherein the multi-branch convolutional neural network model comprises a first branch, a second branch and a third branch, the first branch, the second branch and the third branch respectively comprise an EEGNet convolutional block and a convolutional attention module, and the first branch, the second branch and the third branch are connected through a softmax layer;

the convolution attention module comprises a channel attention sub-module and a space attention sub-module;

the channel attention submodule comprises an average pool layer, a maximum pool layer and a shared network, the shared network comprises a multi-layer perceptron, the multi-layer perceptron is provided with a hidden layer, the average pool layer and the maximum pool layer receive input characteristics and then generate a characteristic map, and the shared network receives the characteristic map and then outputs channel refined characteristics;

calculating activation or output of a jth neuron of an ith filter in the hidden layer by formula 1;

wherein a is _ij Is the ith filter in the hidden layerActivation or output of the jth neuron of the device, f corresponding to the activation function, b _i Is the shared overall bias of the filter, i and K are kernel sizes, W _i ＝[w _i1 ,w _i2 ,…,w _ik ]Is a vector of shared weights, X _j ＝[x _j ,x _j+1 ,…,x _j+k ]Is the vector output by the pre-neuron, T represents the transpose operation;

the spatial attention submodule comprises an average pool layer, a maximum pool layer and a convolution layer, and the channel refined features output refined feature mapping through the average pool layer, the maximum pool layer and the convolution layer;

training the multi-branch convolutional neural network model through the training set;

performing performance evaluation on the trained multi-branch convolutional neural network through the test set to obtain a target multi-branch convolutional neural network model;

and inputting the M I-EEG signals to be classified into the target multi-branch convolutional neural network model to obtain a classification result.

2. The method of classifying M I-EEG signals according to claim 1, wherein the EEGNet convolution block comprises a first convolution layer, a second convolution layer, and a third convolution layer connected in sequence, and wherein the window sizes of the first convolution layer, the second convolution layer, and the third convolution layer are different.

3. The method of classifying MI-EEG signals according to claim 1, wherein said constructing a multi-branch convolutional neural network model, wherein said multi-branch convolutional neural network model comprises a first branch, a second branch and a third branch, said first branch, said second branch and said third branch comprising an EEGNet convolution block and a convolution attention module, respectively, said first branch, said second branch and said third branch being connected by a softmax layer, comprising:

different numbers of parameters are set by EEGNet convolution blocks and convolution attention modules in the first, second and third branches, respectively, to capture different features.

4. A classification system for MI-EEG signals, comprising:

an acquisition module for acquiring an MI-EEG signal dataset;

the preprocessing module is used for preprocessing the M I-EEG signal data set to obtain a preprocessed data set, and dividing a training set and a testing set based on the preprocessed data set;

the construction module is used for constructing a multi-branch convolutional neural network model, wherein the multi-branch convolutional neural network model comprises a first branch, a second branch and a third branch, the first branch, the second branch and the third branch respectively comprise an EEGNet convolutional block and a convolutional attention module, and the first branch, the second branch and the third branch are connected through a softmax layer;

the convolution attention module comprises a channel attention sub-module and a space attention sub-module;

the channel attention submodule comprises an average pool layer, a maximum pool layer and a shared network, the shared network comprises a multi-layer perceptron, the multi-layer perceptron is provided with a hidden layer, the average pool layer and the maximum pool layer receive input characteristics and then generate a characteristic map, and the shared network receives the characteristic map and then outputs channel refined characteristics;

calculating activation or output of a jth neuron of an ith filter in the hidden layer by formula 1;

wherein a is _ij Is the activation or output of the jth neuron of the ith filter in the hidden layer, f corresponds to the activation function, b _i Is the shared overall bias of the filter, i and K are kernel sizes, W _i ＝[w _i1 ,w _i2 ,…,w _ik ]Is a vector of shared weights, X _j ＝[x _j ,x _j+1 ,…,x _j+k ]Is the vector output by the pre-neuron, T represents the transpose operation;

the spatial attention submodule comprises an average pool layer, a maximum pool layer and a convolution layer, and the channel refined features output refined feature mapping through the average pool layer, the maximum pool layer and the convolution layer;

the training module is used for training the multi-branch convolutional neural network model through the training set;

the test module is used for performing performance evaluation on the trained multi-branch convolutional neural network through the test set to obtain a target multi-branch convolutional neural network model;

the multi-branch convolutional neural network model obtains a classification result based on the MI-EEG signals to be classified.

5. The classification system of MI-EEG signal according to claim 4, wherein said EEGNet convolution block comprises a first convolution layer, a second convolution layer and a third convolution layer connected in sequence, and wherein the window sizes of said first convolution layer, said second convolution layer and said third convolution layer are different.

6. The classification system of MI-EEG signal according to claim 4, further comprising a setting module for setting different numbers of parameters for the EEGNet convolution blocks and convolution attention modules in the first, second and third branches, respectively, to capture different features.

7. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the steps of the method according to any one of claims 1 to 3 when the computer program is executed.

8. A non-transitory computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the steps of the method according to any of claims 1 to 3.