Cosine convolution multi-channel group decoding method and system for electroencephalogram classification

Abstract

The present disclosure provides a cosine convolution multi-channel group decoding method and system for electroencephalogram classification, relating to the technical field of artificial intelligence, acquiring different categories of motor imagery electroencephalogram signals to be classified, inputting the motor imagery electroencephalogram signals into an electroencephalogram signal classification model, and outputting classification results of different categories of motor imagery electroencephalogram signals; wherein a cosine convolution time-frequency feature extraction module adopts cosine convolution kernels of different frequencies to decompose and reconstruct different frequency components to obtain multi-scale time-frequency features; the multi-scale time-frequency features are input again into a multi-channel group spatial feature extraction module, each channel subgroup data adopts an independent separable convolution branch network to perform convolution in the channel dimension to obtain time-frequency-space features of each channel subgroup electroencephalogram signal, the time-frequency-space features of each channel subgroup electroencephalogram signal are spliced into a fusion feature vector, and the motor imagery electroencephalogram signal corresponding to the probability of each category of motor imagery is output through a fully connected layer.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the field of artificial intelligence, in particular to a cosine convolution multi-channel group decoding method and system for electroencephalogram classification. BACKGROUND

[0002] The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute prior art.

[0003] Brain-Computer Interface (BCI) technology is a direct connection between the brain and external devices, which converts human thinking activities into signals for controlling external devices or external communication by analyzing collected brain signals, thereby realizing an important way of human-computer interaction. Electroencephalogram (EEG) is widely used in the field of brain-computer interface due to its non-invasiveness, portability, high time resolution, low cost, and suitability for real-time applications.

[0004] Motor Imagery (MI) electroencephalogram (EEG) is a common spontaneous brain signal, which refers to the brain information obtained by imagining one's own limb movement without actual movement. It focuses on subjective initiative, can repair or rebuild damaged motor transmission pathways, and awaken dormant neurons, and has great practical value in active motor rehabilitation.

[0005] Currently, deep learning methods are widely used in motor imagery brain-computer interface, but traditional ordinary convolution has large computational complexity, and the standard linear kernel function used has certain limitations in capturing the nonlinear characteristics of brain signals, making it difficult to fully capture the dynamic change characteristics of periodic electroencephalogram signals. Moreover, existing methods often do not fully utilize the spatial relationship between different electroencephalogram channels when processing multi-channel electroencephalogram data, resulting in insufficient information fusion between electroencephalogram channels and causing information loss. SUMMARY

[0006] To solve the above problems, the present disclosure proposes a cosine convolution multi-channel group decoding method and system for electroencephalogram classification. The first convolution layer is set as a cosine convolution to better extract the time-frequency features of electroencephalogram signals. Multiple branches are used at the position of the second convolution layer to extract the spatial features of different channel group signals. Finally, the extracted time-frequency and spatial features of different channel groups are spliced to realize high-performance motor imagery electroencephalogram signal classification.

[0007] According to some embodiments, the present disclosure adopts the following technical solutions:

[0008] The cosine convolution multi-channel group decoding method for electroencephalogram classification comprises:

[0009] acquire different categories of motor imagery electroencephalogram signals to be classified and pre-process;

[0010] input the pre-processed motor imagery electroencephalogram signals into an electroencephalogram signal classification model to output classification results of the different categories of motor imagery electroencephalogram signals;

[0011] After the motor imagery electroencephalogram signals are input into the electroencephalogram signal classification model, they first enter a cosine convolution time-frequency feature extraction module, decompose and reconstruct different frequency components by using cosine convolution kernels of different frequencies to obtain multi-scale time-frequency features; the multi-scale time-frequency features are input into a multi-channel group space feature extraction module, the multi-scale time-frequency features are grouped into channels according to the number of channel groups, each channel subgroup data is convolved in the channel dimension by using an independent separable convolution branch network to obtain time-frequency space features of each channel subgroup electroencephalogram signal, the time-frequency space features of each channel subgroup electroencephalogram signal are spliced into a fusion feature vector, and the motor imagery electroencephalogram signals are output through a full connection layer to obtain probabilities of each category of motor imagery.

[0012] According to some embodiments, the present disclosure adopts the following technical solutions:

[0013] The cosine convolution multi-channel group decoding system for electroencephalogram classification comprises:

[0014] A data acquisition module is configured to acquire different categories of motor imagery electroencephalogram signals to be classified and pre-process;

[0015] A classification module is configured to input the pre-processed motor imagery electroencephalogram signals into an electroencephalogram signal classification model to output classification results of the different categories of motor imagery electroencephalogram signals;

[0016] After the motor imagery electroencephalogram signals are input into the electroencephalogram signal classification model, they first enter a cosine convolution time-frequency feature extraction module, decompose and reconstruct different frequency components by using cosine convolution kernels of different frequencies to obtain multi-scale time-frequency features; the multi-scale time-frequency features are input into a multi-channel group space feature extraction module, the multi-scale time-frequency features are grouped into channels according to the number of channel groups, each channel subgroup data is convolved in the channel dimension by using an independent separable convolution branch network to obtain time-frequency space features of each channel subgroup electroencephalogram signal, the time-frequency space features of each channel subgroup electroencephalogram signal are spliced into a fusion feature vector, and the motor imagery electroencephalogram signals are output through a full connection layer to obtain probabilities of each category of motor imagery.

[0017] According to some embodiments, the present disclosure adopts the following technical solutions:

[0018] A computer program product comprises a computer program, which, when executed by a processor, implements the cosine convolution multi-channel group decoding method for electroencephalogram classification.

[0019] According to some embodiments, the present disclosure adopts the technical solutions as follows:

[0020] A non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the cosine convolution multi-channel group decoding method for electroencephalogram classification.

[0021] According to some embodiments, the present disclosure adopts the technical solutions as follows:

[0022] An electronic device comprising a processor, a memory, and a computer program; wherein the processor is connected with the memory, and the computer program is stored in the memory; when the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device implements the cosine convolution multi-channel group decoding method for electroencephalogram classification.

[0023] Compared with the prior art, the present disclosure has the beneficial effects as follows:

[0024] The cosine convolution multi-channel group decoding method for electroencephalogram classification of the present disclosure can obtain multi-scale time-frequency features by adopting cosine convolution kernels of different frequencies to decompose and reconstruct different frequency components through a cosine convolution time-frequency feature extraction module, which can better capture the periodic characteristics of electroencephalogram signals by replacing traditional convolution kernels with cosine convolution kernels, effectively extract dynamic characteristics of electroencephalogram signals in time domain and frequency domain, and improve the accuracy of decoding; and the cosine convolution significantly reduces the amount of calculation compared with traditional convolution when calculating, reduces the calculation cost while maintaining high performance, and makes it suitable for real-time application and embedded device implementation.

[0025] The cosine convolution multi-channel group decoding method for electroencephalogram classification of the present disclosure inputs the multi-scale time-frequency features into a multi-channel group spatial feature extraction module again, groups the multi-scale time-frequency features according to the number of channel groups, and convolves each channel subgroup data in the channel dimension by adopting an independent separable convolution branch network, to obtain time-frequency-space features of each channel subgroup electroencephalogram signal; the branch structure of the multi-channel group makes full use of the spatial correlation between channels, overcomes the problem of insufficient spatial information fusion in the existing method when processing multi-channel signals, and improves the generalization ability and robustness of the model. This method can effectively extract frequency domain and cross-channel features in electroencephalogram signals, improve classification performance, and reduce calculation cost. BRIEF DESCRIPTION OF DRAWINGS

[0026] The accompanying drawings, which form a part of this disclosure, are intended to provide further understanding of the present disclosure and are incorporated herein for illustrative purposes. The schematic embodiments of the present disclosure and their descriptions are used to explain the present disclosure, and do not constitute improper limitations on the present disclosure.

[0027] Figure 1 A structural block diagram of a cosine convolution multi-channel group for an embodiment of the present disclosure;

[0028] Figure 2 A classification model structural framework diagram based on a cosine convolution multi-channel group for an embodiment of the present disclosure;

[0029] Figure 3 A flowchart of a cosine convolution multi-channel group decoding method for motor imagery electroencephalogram classification for an embodiment of the present disclosure. DETAILED DESCRIPTION

[0030] The present disclosure will be further described below in conjunction with the accompanying drawings and embodiments.

[0031] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present disclosure. Unless otherwise indicated, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present disclosure belongs.

[0032] It should be noted that the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit the exemplary embodiments according to the present disclosure. As used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and it should also be understood that when the terms "comprise" and / or "include" are used in the specification, there is a presence of a feature, step, operation, device, component, and / or combinations thereof.

[0033] Embodiment 1

[0034] In an embodiment of the present disclosure, a cosine convolution multi-channel group decoding method for electroencephalogram classification is provided, comprising the following steps:

[0035] Step one: obtaining different categories of motor imagery electroencephalogram signals to be classified, and preprocessing;

[0036] Step two: inputting the preprocessed motor imagery electroencephalogram signals into an electroencephalogram signal classification model to output the classification results of different categories of motor imagery electroencephalogram signals;

[0037] Wherein, the motor imagery electroencephalogram signal is input into the electroencephalogram signal classification model, first enters the cosine convolution time-frequency feature extraction module, different frequency components are decomposed and reconstructed by using different frequency cosine convolution kernels, and multi-scale time-frequency features are obtained; the multi-scale time-frequency features are input into the multi-channel group space feature extraction module again, the multi-scale time-frequency features are grouped according to the number of channel groups, each channel subgroup data adopts an independent separable convolution branch network to perform convolution in the channel dimension, and the time-frequency space features of each channel subgroup electroencephalogram signal are obtained; the time-frequency space features of each channel subgroup electroencephalogram signal are spliced into a fusion feature vector, and the probability of each category of motor imagery electroencephalogram signal corresponding to motor imagery is output through a full connection layer.

[0038] As an embodiment, the specific implementation process of the cosine convolution multi-channel group decoding method for electroencephalogram classification of the present disclosure is as follows:

[0039] Step 1: Obtain different categories of motor imagery electroencephalogram signals to be classified, and pre-process;

[0040] Specifically, in order to verify the effectiveness of the present disclosure, the existing public electroencephalogram dataset, the electroencephalogram signal dataset BCI Competition IV dataset IIa of the fourth international brain-computer interface competition in 2008, is used as a reference dataset to compare the classification results. This dataset collects 4s of electroencephalogram data from 4 categories of electroencephalogram signals generated by experimenters imagining left hand, right hand, toes and tongue movement, with a sampling frequency Fs=250Hz. Each sample selects 4s of electroencephalogram data from electroencephalogram channels, that is, the length of the electroencephalogram signal of each sample is , each category collects times, and a total of samples are collected. Among them, 288 samples are taken as a training set, and 288 samples are taken as a test set.

[0041] Step 2: Construct an electroencephalogram signal classification model and train it;

[0042] Specifically, the electroencephalogram signal classification model includes a cosine convolution time-frequency feature extraction module, a multi-channel group space feature extraction module, a feature fusion module and a full connection layer.

[0043] The cosine convolution time-frequency feature extraction module includes a convolution layer with a cosine convolution kernel and a batch normalization layer. The cosine convolution time-frequency feature extraction module uses multiple cosine convolution kernels with different frequencies instead of traditional convolution kernels to filter the electroencephalogram signal, so as to extract multi-scale time-frequency features of the electroencephalogram signal. The input of the cosine convolution time-frequency feature extraction module is the collected electroencephalogram signal , and the input parameters required for cosine convolution are the number of cosine convolution kernels and the length of cosine convolution kernel The calculation formula of cosine convolution kernel is as follows:

[0044]

[0045] Wherein, is a symmetric array centered at zero with a length of , that is, {-4, -3, -2, -1, 0, 1, 2, 3, 4}, represents the amplitude of the j th trainable cosine convolution kernel, which is initialized as a random quantity obeying Gaussian distribution with mean 0 and standard deviation 1, and through iterative updating in the training process, it can amplify or suppress the signal components of different frequency bands, so as to better adapt to the periodic characteristics in the electroencephalogram signal; represents the frequency parameter, through which the division of different frequency bands is realized, The initialization is as follows:

[0046]

[0047] Wherein represents the starting frequency of the set electroencephalogram data, represents the cut-off frequency of the set electroencephalogram data, represents the number of cosine convolution filter kernels.

[0048] Wherein, the cosine convolution determines the cut-off frequency band of the decoded electroencephalogram signal by giving and The frequency parameter is fixed after initialization and does not update with training, and the fixed value can determine the extraction frequency band of the corresponding convolution kernel. In this embodiment, the extraction frequency bands of the corresponding convolution kernels are [0 Hz, 5 Hz], [5 Hz, 10 Hz], [10 Hz, 15 Hz], [15 Hz, 20 Hz], [20 Hz, 25 Hz], [25 Hz, 30 Hz], [30 Hz, 35 Hz], and [35 Hz, 40 Hz].

[0049] Put the cosine convolution kernel into the two-dimensional convolution to obtain the multi-scale time-frequency features of the cosine convolution output :

[0050]

[0051] Wherein, represents the cosine convolution kernel, * represents the convolution operation, represents the input data, represents thej the amplitude of the trainable cosine convolution kernel, representing a frequency parameter.

[0052] The cosine convolution filters the electroencephalogram signals in different frequency bands by adopting different frequency parameters, thereby obtaining multi-scale time-frequency features. The features can better capture the frequency information contained in the signals and improve the classification effect. The output of the cosine convolution is subjected to batch normalization to obtain data , so that the data distribution is more stable, the risk of gradient disappearance and gradient explosion is reduced, and the robustness of the model is enhanced.

[0053] Batch normalization is a regularization technique aimed at accelerating the training process of a neural network and improving the stability of the model. It standardizes the input features of each mini-batch to a distribution with a mean of 0 and a standard deviation of 1, and introduces learnable parameters for scaling and shifting to preserve the expressiveness of the model. The improvement of the present disclosure is that the cosine convolution realizes the extraction of time-frequency features in different frequency bands of the same signal through frequency parameters , which are fixed parameters set according to a certain rule and cannot be trained.

[0054] Further, a multi-channel group spatial feature extraction module is constructed, which includes a channel grouping operation and a depth separable convolution layer in each group.

[0055] The input of the multi-channel group spatial feature extraction module is the multi-scale time-frequency features of the electroencephalogram signals output by the cosine convolution time-frequency feature extraction module . The multi-channel group spatial feature extraction module divides the input multi-channel electroencephalogram signals into groups according to certain channel correlation. The number of channels and the position of the channels in each group are different. The input multi-scale time-frequency features are divided into 10 channel subgroups according to the total number of channels, a certain step size, and different starting positions. The signals in each channel subgroup have strong spatial correlation. The electroencephalogram signals in each subgroup correspond to a branch of the network, and a depth separable convolution is present in each branch to extract the spatial features of the electroencephalogram signals in the subgroup. Channel grouping is a grouping obtained by interval sampling of the electroencephalogram channels according to a certain rule, which can enhance the spatial feature extraction capability of the model for the electroencephalogram signals. The ordering of all channels of the electroencephalogram signals starts from 0, and the channel grouping is generated according to the following rules:

[0056] 1. The first four channel groups take 0 as the starting position, and generate channel indexes with a step size of 1, 2, 3, and 4, respectively. The four channel groups are selected by cyclic jumping according to the channel position, and the channel indexes in each group are separated by 1, 2, 3, and 4 positions, respectively.

[0057] When the step size is 1, the channel group with 0 as the starting position contains the indexes of ;

[0058] When the step size is 2, the channel group with 0 as the starting position contains the indexes of ;

[0059] When the step size is 3, the channel group with 0 as the starting position contains the indexes of ;

[0060] When the step size is 4, the channel group with 0 as the starting position contains the indexes of .

[0061] 2. The last three channel groups generate channel indexes with 1 as the starting position, with step sizes of 2, 3, and 4 respectively, and the three channel groups are selected according to the cyclic jump of the channel positions, and the channel indexes in each group are separated by 2, 3, and 4 positions respectively.

[0062] When the step size is 2, the channel group with 1 as the starting position contains the indexes of ;

[0063] When the step size is 3, the channel group with 1 as the starting position contains the indexes of ;

[0064] When the step size is 4, the channel group with 1 as the starting position contains the indexes of .

[0065] 3. The last two channel groups generate channel indexes with 2 as the starting position, with step sizes of 3 and 4 respectively, and the two channel groups are selected according to the cyclic jump of the channel positions, and the channel indexes in each group are separated by 3 and 4 positions respectively.

[0066] When the step size is 3, the channel group with 2 as the starting position contains the indexes of ;

[0067] When the step size is 4, the channel group with 2 as the starting position contains the indexes of .

[0068] 4. The last channel group generates channel indexes with 3 as the starting position, with a step size of 4, and the channel group is selected according to the cyclic jump of the channel positions, and the channel indexes in each group are separated by 4 positions.

[0069] When the step size is 4, the channel group with 2 as the starting position contains the indexes of .

[0070] For each channel sub-group data after the channel grouping ( The number of channels representing the i-th channel group) is convolved in the channel dimension with an independent deep separable convolution branch network in the branch to extract the spatial features of the electroencephalogram signals in the subgroup The number of convolution kernels in each branch of the multi-channel group spatial feature extraction module is represented by K.

[0071] Wherein, the depth separable convolution further optimizes the convolution process, first performs a separate convolution operation (depth convolution) on each input channel, and then performs point-by-point convolution to mix channel information.

[0072] Further, a feature fusion and classification module is constructed, and the input data of the feature fusion and classification module is the time-frequency-space features of the multiple groups of electroencephalogram signals output by the multi-channel group spatial feature extraction module In the dimension of the original electroencephalogram channel, the input features are spliced as follows to obtain data :

[0073]

[0074] Wherein, The output of a channel group of the multi-channel group spatial feature extraction module is represented by Y, and T represents the length of the electroencephalogram signal. The number of convolution kernels in each branch of the multi-channel group spatial feature extraction module is represented by K.

[0075] The fused feature map is applied to the ELU activation function to introduce nonlinearity, and then further convolution, batch normalization and average pooling operations are performed to gradually reduce the size of the output feature map. Finally, the feature map is converted to a one-dimensional feature vector through the flattening layer, then connected to the fully connected layer, and the activation function Softmax is used to output the probability of each class corresponding to the electroencephalogram data.

[0076] Wherein, the average pooling is used for dimension reduction operation, which is used for feature extraction in deep learning model. It reduces the spatial dimension (height and width) of the input feature map by calculating the average value of elements within the local receptive field, thereby reducing the amount of calculation and parameters, while retaining the global information of the input features.

[0077] ELU is a kind of nonlinear activation function, which is used in the hidden layer to solve the problem of gradient disappearance and improve the expression ability of the model to negative features; Softmax is a kind of activation function, which is used in the output layer of multi-classification task to convert the output value into probability distribution, helping the model to determine the classification result.

[0078] Further, according to the output classification result, the corresponding control command is converted to control the external device.

[0079] To illustrate the effectiveness of the present disclosure in decoding motor imagery electroencephalogram signals, the classification accuracy is used as an evaluation index based on the obtained category labels:

[0080]

[0081] where TP is the number of true positives, TN is the number of true negatives, FP is the number of false positives, and FN is the number of false negatives.

[0082] Comparing the present disclosure with the EEGNet lightweight network dedicated to electroencephalogram decoding and the ablation results of each module, it is found that under the same experimental conditions, the motor imagery electroencephalogram classification accuracy and the average accuracy of the present disclosure are higher than those of other groups, which shows that the cosine convolution module and the multi-channel group module have a synergistic complementary effect, and the combination of the two enhances the model's ability to extract frequency domain features and spatial features, verifying the effectiveness of the present disclosure in decoding electroencephalogram signals, and laying a certain foundation for converting the category labels obtained by decoding electroencephalogram signals into control instructions to accurately control external devices.

[0083] Table 1: Comparison of intra-group results of the present disclosure

[0084]

[0085] The results in Table 1 show that the cosine convolution multi-channel group decoding method improves the accuracy by about 3% compared to the traditional EEGNet, which shows the effectiveness of the present invention in improving model performance, and the combination of cosine convolution and multi-channel group enhances the traditional model's ability to extract time-frequency-space features of electroencephalogram signals. The present disclosure overcomes the problem of insufficient spatial information fusion in existing methods when processing multi-channel signals, improves the generalization ability and robustness of the model, and effectively extracts frequency domain and cross-channel features in electroencephalogram signals, improves classification performance, and reduces computational cost.

[0086] Embodiment 2

[0087] In an embodiment of the present disclosure, a cosine convolution multi-channel group decoding system for electroencephalogram classification is provided, comprising:

[0088] A data acquisition module for acquiring different categories of motor imagery electroencephalogram signals to be classified and preprocessing;

[0089] A classification module for inputting the preprocessed motor imagery electroencephalogram signals into an electroencephalogram signal classification model and outputting the classification results of different categories of motor imagery electroencephalogram signals;

[0090] Wherein, the motor imagination electroencephalogram signal is input to the electroencephalogram signal classification model, first enters the cosine convolution time-frequency feature extraction module, different frequency cosine convolution kernels are used to decompose and reconstruct different frequency components, and multi-scale time-frequency features are obtained; the multi-scale time-frequency features are input to the multi-channel group space feature extraction module again, the multi-scale time-frequency features are grouped according to the number of channel groups, each channel subgroup data adopts an independent separable convolution branch network to perform convolution in the channel dimension, and the time-frequency space features of each channel subgroup electroencephalogram signal are obtained; the time-frequency space features of each channel subgroup electroencephalogram signal are spliced into a fusion feature vector, and the motor imagination electroencephalogram signal corresponding to each category of motor imagination is output through a full connection layer.

[0091] As an embodiment, the data acquisition module includes a data acquisition device and an A / D converter, and each module of the classification model is embedded in a cosine convolution time domain filtering operation unit, a convolution space filtering operation unit and a full connection layer operation unit in a high-performance computer to perform signal processing and classification, obtain the classification label of the sample, and convert it into a control instruction for an external device.

[0092] Embodiment 3

[0093] An embodiment of the present disclosure provides a computer program product, comprising a computer program, wherein the computer program is executed by a processor to realize the cosine convolution multi-channel group decoding method for electroencephalogram classification.

[0094] Embodiment 4

[0095] An embodiment of the present disclosure provides a non-transitory computer readable storage medium for storing computer instructions, wherein the computer instructions are executed by a processor to realize the cosine convolution multi-channel group decoding method for electroencephalogram classification.

[0096] Embodiment 5

[0097] An embodiment of the present disclosure provides an electronic device, comprising a processor, a memory and a computer program; wherein the processor is connected with the memory, and the computer program is stored in the memory; when the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device executes the cosine convolution multi-channel group decoding method for electroencephalogram classification.

[0098] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 one or more flow or blocks

[0099] These computer program instructions can also be loaded into a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 one or more flow or blocks

[0100] Although the present disclosure has been described with reference to specific implementations, it will be apparent to those skilled in the art that various modifications and changes can be made thereto without departing from the scope of the present disclosure.

Claims

1. A cosine convolution multi-channel group decoding method for electroencephalographic classification, characterized in that, The method comprises the following steps: acquiring different categories of motor imagery electroencephalogram signals to be classified and preprocessing; inputting the preprocessed motor imagery electroencephalogram signals into an electroencephalogram signal classification model to output classification results of the different categories of motor imagery electroencephalogram signals; wherein after the motor imagery electroencephalogram signals are input into the electroencephalogram signal classification model, first, the motor imagery electroencephalogram signals enter a cosine convolution time-frequency feature extraction module, different frequency components are decomposed and reconstructed by using cosine convolution kernels with different frequencies, and multi-scale time-frequency features are obtained; the multi-scale time-frequency features are input into a multi-channel group spatial feature extraction module again, the multi-scale time-frequency features are grouped according to the number of channel groups, and each channel subgroup data is convolved in the channel dimension by using an independent separable convolution branch network to obtain time-frequency spatial features of each channel subgroup electroencephalogram signal; the time-frequency spatial features of each channel subgroup electroencephalogram signal are spliced into a fusion feature vector, and a full connection layer is used to output probabilities of each category of motor imagery corresponding to the motor imagery electroencephalogram signals; The electroencephalogram signal classification model comprises a cosine convolution time-frequency feature extraction module, a multi-channel group space feature extraction module, a feature fusion module and a full connection layer. n and the length of the cosine convolution kernel L The calculation formula of the cosine convolution kernel is as follows: ; wherein, is a symmetric array with zero as the center and the length of , and denotes the amplitude of the jth trainable cosine convolution kernel. wherein, denotes a frequency parameter, is initialized as follows: ; wherein, denotes a start frequency of the set EEG data, denotes a stop frequency of the set EEG data, the frequency parameter is fixed after initialization and is not updated during training, the fixed value determines the extraction frequency band of the corresponding convolution kernel; wherein the input of the multi-channel group spatial feature extraction module is the multi-scale time-frequency features of the motor imagery electroencephalogram signals, the input multi-scale time-frequency features are divided into multiple channel subgroups according to a grouping rule in accordance with a set step and a starting position, so that the signals in each channel subgroup have strong spatial correlation, and the ordering of all channels of the electroencephalogram signal starts from 0.

2. The cosine convolution multi-channel group decoding method for electroencephalographic classification of claim 1, wherein, The different categories of motor imagery electroencephalogram signals to be classified are acquired, and the preprocessing includes: collecting the motor imagery electroencephalogram signals by using an electroencephalograph and an A / D converter, wherein the collected motor imagery electroencephalogram signals contain A sampling frequency is set for different categories, M times are collected for each category, and the total number of the collected motor imagery electroencephalogram signals is The collected motor imagery electroencephalogram signals are subjected to denoising filtering and segmentation preprocessing.

3. The cosine convolution multi-channel group decoding method for electroencephalographic classification of claim 1, wherein, By using different frequency cosine convolution kernels to decompose and reconstruct different frequency components, multi-scale frequency domain features are obtained. By putting the cosine convolution kernel into two-dimensional convolution, the cosine convolution output is obtained as: ; wherein, represents the cosine convolution kernel, * represents the convolution operation, represents the input data, represents the amplitude of the first j trainable cosine convolution kernel, represents the frequency parameter.

4. The cosine convolution multi-channel group decoding method for electroencephalographic classification of claim 1, wherein, The output multiple groups of time-frequency spatial features are spliced to obtain a fused feature map, the fused feature map is subjected to further convolution and pooling operations, the size of the feature map is gradually reduced, and finally the feature map is converted into a one-dimensional fusion feature vector by using a flattening layer, then a full connection layer is connected, an activation function is used, and probabilities of each category of motor imagery corresponding to the electroencephalogram data are output.

5. A cosine convolution multi-channel group decoding system for electroencephalographic classification, characterized by, The method comprises the following steps: a data acquisition module is configured to acquire different categories of motor imagery electroencephalogram signals to be classified and preprocess the motor imagery electroencephalogram signals; a classification module is configured to input the preprocessed motor imagery electroencephalogram signals into an electroencephalogram signal classification model to output classification results of the different categories of motor imagery electroencephalogram signals; wherein after the motor imagery electroencephalogram signals are input into the electroencephalogram signal classification model, first, the motor imagery electroencephalogram signals enter a cosine convolution time-frequency feature extraction module, different frequency components are decomposed and reconstructed by using cosine convolution kernels with different frequencies, and multi-scale time-frequency features are obtained; the multi-scale time-frequency features are input into a multi-channel group spatial feature extraction module again, the multi-scale time-frequency features are grouped according to the number of channel groups, and each channel subgroup data is convolved in the channel dimension by using an independent separable convolution branch network to obtain time-frequency spatial features of each channel subgroup electroencephalogram signal; the time-frequency spatial features of each channel subgroup electroencephalogram signal are spliced into a fusion feature vector, and a full connection layer is used to output probabilities of each category of motor imagery corresponding to the motor imagery electroencephalogram signals; The EEG signal classification model includes a cosine convolution time-frequency feature extraction module, a multi-channel group spatial feature extraction module, a feature fusion module, and a fully connected layer. The cosine convolution time-frequency feature extraction module uses cosine convolution kernels instead of traditional convolution kernels, and the input parameter required for cosine convolution is the number of cosine convolution kernels. n and the length of the cosine convolution kernel L The formula for calculating the cosine convolution kernel is as follows: ;in, It is a zero-centered structure with a length of Symmetric arrays, This represents the magnitude of the j-th trainable cosine convolution kernel; wherein, denotes a frequency parameter, is initialized as follows: ; wherein, denotes a start frequency of the set of EEG data, denotes a stop frequency of the set of EEG data, the frequency parameter is fixed after initialization and is not updated during training, the fixed value determining the extracted frequency band of the corresponding convolution kernel; wherein the input of the multi-channel group spatial feature extraction module is the multi-scale time-frequency features of the motor imagery electroencephalogram signals, the input multi-scale time-frequency features are divided into multiple channel subgroups according to a grouping rule in accordance with a set step and a starting position, so that the signals in each channel subgroup have strong spatial correlation, and the ordering of all channels of the electroencephalogram signal starts from 0.

6. A computer program product comprising a computer program, characterized in that, The computer program, when executed by a processor, implements the cosine convolution multi-channel group decoding method for electroencephalogram classification according to any one of claims 1-4.

7. A non-transitory computer-readable storage medium, comprising: The non-transitory computer readable storage medium is configured to store computer instructions, and the computer instructions, when executed by a processor, implement the cosine convolution multi-channel group decoding method for electroencephalogram classification according to any one of claims 1-4.

8. An electronic device, comprising: Comprise: A processor, a memory and a computer program; wherein the processor is connected with the memory, and the computer program is stored in the memory; when the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device executes the cosine convolution multi-channel group decoding method for electroencephalogram classification according to any one of claims 1-4.

Images