EEG channel selection method and system based on double-branch collaborative guide learning

Abstract

The invention discloses an EEG channel selection method and system based on double-branch collaborative guide learning. The method comprises the steps that original electroencephalogram signals are collected; the original electroencephalogram signals are input into the channel importance evaluation module through the signal input module, the smooth score of each channel is calculated, and a set number of optimal channels are selected; the signal feature extraction module extracts the space-time features of the optimal channel and pools and projects the space-time features to obtain final embedded features; the embedded features are input into the double branches, global average pooling is carried out on the embedded features, and prediction label probability distribution is obtained through multi-class prediction; reconstructing the complete electroencephalogram signals according to the embedded features to obtain reconstructed electroencephalogram signals; constructing a total loss function, minimizing the total loss, and adjusting the weight of the electroencephalogram channel selection model; and performing channel selection by using the trained electroencephalogram channel selection model. According to the method, through a collaborative learning mechanism, on the basis of ensuring the performance, the number of channels required by electroencephalogram signals is remarkably reduced.

Description

Technical Field

[0001] This invention relates to the field of electroencephalogram (EEG) signal processing technology, and in particular to an EEG channel selection method and system based on bi-branch collaborative guided learning. Background Technology

[0002] Electroencephalography (EEG) signals are important bioelectrical signals widely used in neuroscience, brain-computer interfaces, neurological disease diagnosis, rehabilitation engineering, and cognitive research. Traditional EEG acquisition systems typically use high-density electrode arrays (such as 64, 128, or 256 channels) to capture as comprehensive neural activity as possible. These systems record the potential changes of neuronal clusters in different brain regions by placing multiple electrodes on the scalp surface, thus generating multi-channel, high-temporal-resolution electrophysiological data. However, this high-density acquisition method has the following main problems:

[0003] (a) High cost and complex operation: High-density electrode placement requires professional installation and debugging, and 30-60 minutes of preparation work is required before each collection, including electrode positioning, conductive gel injection, impedance testing, etc., which is time-consuming and easily affected by the environment and the operator's skill level. In addition, high-density electrode equipment is expensive, with the cost of a single system reaching hundreds of thousands of yuan, increasing the threshold for research and clinical application;

[0004] (b) Data redundancy and heavy computational burden: Significant information redundancy exists in a large number of channels. Adjacent electrodes often capture similar neural signals, leading to dimensionality explosion and wasted computational resources. For example, a 128-channel system with a sampling rate of 1000Hz will generate 7.68MB of raw data per minute. Long-term recording will generate massive amounts of data, posing huge challenges to storage, transmission and real-time processing.

[0005] (c) Lack of task relevance in channel selection strategies: Existing channel selection methods are mostly based on statistical metrics (such as signal-to-noise ratio and mutual information) or heuristic algorithms (such as genetic algorithms and particle swarm optimization), making it difficult to simultaneously consider signal integrity and specific task performance. These methods are usually static and task-detached, and cannot adaptively adjust the channel selection strategy according to task requirements, resulting in unstable performance in different application scenarios.

[0006] (d) Poor wearing comfort: High-density electrode caps are inconvenient to wear and can cause discomfort to subjects after prolonged use, which limits the application prospects of EEG technology in daily monitoring and long-term wearable devices. Summary of the Invention

[0007] This invention provides an EEG channel selection method and system based on dual-branch collaborative guided learning to solve the technical problems mentioned in the background.

[0008] To achieve the above object, the technical scheme of the present application is as follows:

[0009] The present application provides an EEG channel selection method based on double-branch collaborative guidance learning, comprising the following steps:

[0010] S1, the original EEG signal EEG is collected, and then an EEG channel selection model is constructed, the EEG channel selection model comprising a signal input module, a channel importance evaluation module, a signal feature extraction module, a collaborative guidance learning module and a channel selection output module connected in sequence;

[0011] S2, the original EEG signal EEG is input into the channel importance evaluation module through the signal input module, the channel importance evaluation module calculates the smooth score of each channel, and a certain number of optimal channels are selected according to the smooth score of each channel;

[0012] S3, the selected optimal channels are input into the signal feature extraction module, the spatiotemporal characteristics of the optimal channels are extracted, and after pooling and projection, the final embedding features are obtained;

[0013] S4, the embedding features are input into the task-aware branch and the task-agnostic branch in the collaborative guidance learning module, in the task-aware branch, the embedding features are globally averaged and pooled to obtain the predicted label probability distribution through multi-class prediction; in the task-agnostic branch, the embedding features are used to reconstruct the complete EEG signal to obtain the reconstructed EEG signal;

[0014] S5, a total loss function is constructed according to the predicted label probability distribution and the reconstructed EEG signal, S2 to S4 are iterated, the total loss is minimized, and the weights of the EEG channel selection model are adjusted during the iteration process to obtain the trained EEG channel selection model;

[0015] S6, the trained EEG channel selection model is deployed to the device end, the channels in the actual EEG signal EEG are selected by using the trained EEG channel selection model, a channel selection matrix is obtained, and the selection of the EEG channel is completed.

[0016] The present application further provides an EEG channel selection system based on double-branch collaborative guidance learning, comprising a device end, the device end selects the EEG channel by using the EEG channel selection method, and the device end comprises:

[0017] An EEG signal acquisition unit is configured to receive the original EEG signal EEG from the electrode and pre-process the original EEG signal EEG;

[0018] The signal processing unit is connected with the electroencephalogram signal acquisition unit, and a trained electroencephalogram channel selection model is built-in, and the signal processing unit realizes real-time calculation function based on an embedded processor or a field programmable gate array (FPGA);

[0019] The man-machine interaction unit is electrically connected with the signal processing unit, and is used for providing a touch screen or a button interface for an operator to control and view results.

[0020] The present application has the following beneficial effects:

[0021] 1. The present application discloses an EEG channel selection method based on double-branch collaborative guidance learning, which internally uses an electroencephalogram channel selection model. The electroencephalogram channel selection model is built-in with a collaborative guidance learning module, which adopts double-branch learning training to improve the learning efficiency of the electroencephalogram channel selection model and achieve better training effect. In addition, the collaborative learning mechanism in the collaborative guidance learning module significantly reduces the number of channels required for the electroencephalogram signal, i.e., reduces the number of electrodes, on the basis of ensuring performance. Therefore, the present application also reduces the operation complexity caused by the device cost and electrode wearing, and improves the wearing comfort of the electrode cap.

[0022] In addition, the EEG channel selection method disclosed by the present application has strong adaptability and can be applied to various downstream tasks such as emotion recognition, intention recognition, medical diagnosis, etc.

[0023] 2. The present application also discloses an EEG channel selection system based on double-branch collaborative guidance learning, which is built-in with a man-machine interaction unit. The man-machine interaction unit enables the present application to have channel selection visualization capability, which is convenient for neuroscientific research and analysis. DETAILED DESCRIPTION

[0024] Figure 1 The flowchart of the EEG channel selection method in the present application. DETAILED DESCRIPTION

[0025] In order to facilitate the understanding of the present application, the present application will be described more fully below with reference to the related drawings. The preferred embodiments of the present application are shown in the drawings. However, the present application can be realized in many other different forms, and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present application more thorough and comprehensive.

[0026] In addition, the terms "first", "second" are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined as "first", "second" can explicitly or implicitly include one or more features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0027] Referring to Figure 1 The embodiment of the application provides an EEG channel selection method based on double-branch cooperative guidance learning, comprising the following steps:

[0028] S1, original high-density electroencephalogram (EEG) signals are collected, and an electroencephalogram channel selection model is constructed, the electroencephalogram channel selection model comprising a signal input module, a channel importance evaluation module, a signal feature extraction module, a cooperative guidance learning module and a channel selection output module connected in sequence;

[0029] S2, the original high-density electroencephalogram (EEG) signals are input into the channel importance evaluation module through the signal input module, the channel importance evaluation module calculates the smooth score of each channel, and a set number of optimal channels are selected according to the smooth score of each channel;

[0030] S3, the selected optimal channels are input into the signal feature extraction module, the spatiotemporal characteristics of the optimal channels are extracted, and after pooling and projection, the final embedding features are obtained;

[0031] S4, the embedding features are input into the task-aware branch and the task-agnostic branch in the cooperative guidance learning module, in the task-aware branch, the embedding features are globally averaged and pooled to obtain a predicted label probability distribution through multi-class prediction; in the task-agnostic branch, the embedding features are used to reconstruct the complete high-density electroencephalogram (EEG) signals to obtain reconstructed high-density electroencephalogram (EEG) signals;

[0032] S5, a total loss function is constructed according to the predicted label probability distribution and the reconstructed high-density electroencephalogram (EEG) signals, S2 to S4 are iterated, the total loss is minimized, and the weights of the electroencephalogram channel selection model are adjusted during the iteration process to obtain a trained electroencephalogram channel selection model;

[0033] S6, the trained electroencephalogram channel selection model is deployed to a device end, the channels in the actual high-density electroencephalogram (EEG) signals are selected using the trained electroencephalogram channel selection model, a channel selection matrix is obtained, and the selection of the electroencephalogram channels is completed. For example, 8-16 key channels are selected from 64 channels, and the selection of the channels is the selection of the electrodes.

[0034] The EEG channel selection model internally used by the application is provided with a cooperative guidance learning module, the cooperative guidance learning module adopts double branches for learning and training, thereby improving the learning efficiency of the EEG channel selection model and achieving better training effect; in addition, the cooperative guidance learning module is provided with a cooperative learning mechanism, which significantly reduces the number of channels required for high-density EEG signals on the basis of ensuring performance, that is, reduces the number of electrodes, so that the application also reduces the operation complexity caused by the electrode wearing and improves the wearing comfort of the electrode cap.

[0035] In addition, the EEG channel selection method disclosed by the application has strong adaptability and can be applied to various downstream tasks such as emotion recognition, intention recognition, medical diagnosis and the like.

[0036] In some embodiments, the channel importance evaluation module comprises a one-dimensional convolutional neural network Conv1D, a global average pooling layer and a multilayer perceptron MLP connected in sequence, wherein the multilayer perceptron MLP comprises a fully connected layer and a rectified linear unit activation function connected in sequence.

[0037] The signal feature extraction module comprises a spatio-temporal convolution network ConvEEG, a linear unit activation function ELU, a pooling layer and a projection layer connected in sequence.

[0038] The cooperative guidance learning module comprises two branches of a task-aware branch and a task-agnostic branch connected in parallel, wherein the task-aware branch comprises a global average pooling layer and a classifier connected in sequence; the task-agnostic branch comprises an unembedding network and a channel super-resolution network connected in sequence; wherein the unembedding network comprises a transpose convolution layer and an up-sampling layer connected in sequence; the channel super-resolution network comprises a graph convolution network and a feedforward network connected in parallel.

[0039] In some embodiments, the S2 specifically comprises the following steps:

[0040] S21, inputting the original high-density electroencephalogram EEG to the channel importance evaluation module through a signal input module, modeling the information of each channel by a one-dimensional convolutional neural network Conv1D and a multilayer perceptron MLP in the channel importance evaluation module, and calculating the importance score of each channel; the formula is as follows:

[0041]

[0042] Then, according to the importance score of each channel, the smooth score of each channel is calculated.

[0043] S22, selecting C l optimal channels according to the smooth score of each channel and using a differentiable Gumbel-TopK mechanism.

[0044] In some embodiments, the S21 specifically comprises the following steps:

[0045] S211, inputting the original high-density electroencephalogram (EEG) signal to the channel importance evaluation module through the signal input module, and performing feature extraction on the time sequence of each channel by a one-dimensional convolutional neural network (Conv1D) in the channel importance evaluation module to obtain time features, specifically as follows:

[0046] Z = Conv1D(X h , k t , d, N f )

[0047] wherein Z represents the time features, and R represents a real set, B (batch size) represents a batch size, C h represents the total number of channels used in the electroencephalogram acquisition, and the total number of channels C h is usually 32, 64, 128 or 256, i.e. the number of electrodes; T' represents the dimension of the time features; X h represents the original high-density electroencephalogram (EEG) signal, and T represents the length of the time sequence of each channel, i.e. the number of sampling points, which is usually several hundred to several thousand sampling points; k t is the size of the time convolution kernel (such as 16); d is the convolution step (such as 2); N f is the number of convolution filters (such as 32); and the time sequence length T and the dimension T' of the time features satisfy the following relationship:

[0048]

[0049] S212, compressing the dimension of the time features Z by a global average pooling layer to obtain compressed time features Z', specifically as follows:

[0050]

[0051] wherein GlobalAvgPool represents the dimension compression operation of the global average pooling layer;

[0052] S213, calculating the importance score of each channel according to the compressed time features Z' and using a multi-layer perceptron (MLP), specifically as follows:

[0053]

[0054] wherein f represents the importance score of the channel, FC1 and FC2 are respectively dimension C h ×64 and 64×C ha fully connected layer; ReLU is a rectified linear unit activation function;

[0055] S214, in order to enhance the stability of channel selection, a history memory mechanism is introduced, and the importance scores of each channel are respectively smoothed to obtain the smoothed scores of each channel, as follows:

[0056]

[0057] wherein, represents the smoothed score of the channel, γ represents the history influence factor (typical value is 0.1), h b-1 represents the history channel selection frequency of the b-1th round; and h b-1 satisfies the following relationship:

[0058]

[0059] wherein, h b is the history channel selection frequency of the bth round; α is the momentum update coefficient (typical value is 0.9); S b is the channel selection matrix of the bth round, represents the expected operation, and the expected operation is to take the average of the batch dimension.

[0060] In some embodiments, the S22 specifically comprises the following steps:

[0061] S221, add Gumbel noise to the smoothed score of each channel and perform softmax operation to obtain the channel selection probability p b , as follows:

[0062]

[0063] wherein, g b =-log(-log(u b )) is Gumbel noise; u b ~Uniform(0, 1) is a random variable sampled from a uniform distribution; represents the smoothed score of the bth batch; τ is a temperature parameter to control the randomness of sampling, τ is adjusted by an exponential annealing strategy, and the calculation formula of τ is:

[0064]

[0065] wherein, τ init is the initial temperature (such as 1.0); β is the annealing rate (such as 1000); t is the current training step number; exp represents the exponential function with the natural constant e as the base;

[0066] S222, the selection probability pb Perform a TopK operation to obtain the index of each channel, as follows:

[0067] I b =TopK(p b C l )

[0068] in, The index of the selected channel;

[0069] S223. Select the top C channels with the highest probabilities based on the index of each channel. l Each channel yields a selected high-density EEG signal, as follows:

[0070] X l [b, i, :] = X h [b,I b [i],:]

[0071] in, This indicates that only the high-density EEG signals ranked by importance (C) were retained after selection. l Bit-level channel data.

[0072] In some embodiments, S3 specifically includes the following steps:

[0073] S31, Select the high-density EEG signal X l The input is fed into the signal feature extraction module, where the spatiotemporal convolutional network ConvEEG processes the selected high-density EEG signal X. l Perform temporal convolution to obtain local temporal features. Specifically as follows:

[0074]

[0075] Where, k t The size of the convolution kernel over time; For learnable temporal convolutional kernel weights, c represents the channel number, and F represents the number of temporal filters (e.g., 32 or 64); f b This indicates the filter number, and t is the time index;

[0076] S32. Perform spatial convolution along the channel dimension, and apply it based on the local temporal feature X. temp By fusing spatial information from different channels, we can obtain the fused spatial features. Specifically as follows:

[0077]

[0078] in, Spatial convolution kernel;

[0079] S33, compressing the fused spatial features X by an exponential linear unit activation function ELU and average pooling spatial to obtain the pooled features Specifically, the following applies:

[0080]

[0081] where k is the size of the pooling kernel (e.g., 4 or 8); t' is the time index after pooling; p

[0082] S34, projecting the pooled features X pooled to an embedding space by a 1x1 convolution to obtain the final embedding features Specifically, the following applies:

[0083]

[0084] where e is the embedding dimension index, is the projection matrix, and E is the embedding dimension (e.g., 128 or 256).

[0085] In some embodiments, the S4 specifically includes the following steps:

[0086] S41, inputting the embedding features X patch into a task-aware branch and a task-agnostic branch of a co-guided learning module, if the task-aware branch, then entering S42, otherwise entering S43;

[0087] S42, in the task-aware branch, first performing global average pooling on the embedding features X patch and then inputting the pooled features into a classifier to perform multi-class prediction to obtain a predicted label probability distribution

[0088] S43, in the task-agnostic branch, first mapping the embedding features X patch back to a time series by an unembedding network to obtain a mapped back high-density electroencephalogram signal reconstructing a complete high-density electroencephalogram signal from the mapped back high-density electroencephalogram signal to obtain a reconstructed high-density electroencephalogram signal

[0089] In some embodiments, the entire process of the S42 is represented by the following formula:

[0090]

[0091] ​​where g denotes the pooled embedding features; Dropout denotes the dropout layer; softmax denotes the normalized exponential function; FC3 and FC4 are fully connected layers with dimensions E x 128 and 128 x K, respectively; K is the number of classes, to predict the label probability distribution;

[0092] the mapped high-density electroencephalogram signal in S43 The formula is as follows:

[0093]

[0094] where DeConvEEG is the transposed version of the spatial-temporal convolution network ConvEEG, including transposed convolution, up-sampling, etc.

[0095] the reconstructed high-density electroencephalogram signal in S43 The formula is as follows:

[0096]

[0097] where SuperResolution is the channel super-resolution network. The channel super-resolution network uses the graph convolution network GCN to capture the topological relationship between channels, and uses the feedforward network for channel expansion, so It can also be represented by the following formula:

[0098]

[0099] where, is the channel connection matrix (predefined based on the physical position relationship of the electrodes).

[0100] In some embodiments, the total loss function in S5 is as follows:

[0101]

[0102] where, denotes the total loss; denotes the reconstruction loss, used to maintain signal information; denotes the classification loss, used to ensure task performance; denotes the diversity loss, used to encourage channel selection diversity; λ class , λ div denote the hyperparameters of the classification loss and the diversity loss, respectively, used to adjust the weights between different loss terms, and the typical values are 1.0 and 0.1, respectively;

[0103] The calculation formula of the reconstruction loss is as follows:

[0104]

[0105] where ||·||2 is L2 norm, used to measure the mean square error between the original electroencephalogram EEG and the reconstructed electroencephalogram;

[0106] the calculation formula of the classification loss is:

[0107]

[0108] where y b，k is the one-hot encoding of the true label of the kth class of the bth batch, is the probability of the predicted label of the kth class of the bth batch, and K is the number of classes;

[0109] the calculation formula of the diversity loss is:

[0110]

[0111] where s ：，c represents the selection probability of the cth channel in different samples, and std(·) is the standard deviation.

[0112] In order to facilitate understanding of the EEG channel selection method in the present application, the following is illustrated by way of example:

[0113] Step S1: High-density electroencephalogram signal acquisition. In this embodiment, the SEED dataset is used as the original high-density electroencephalogram EEG. The SEED dataset contains high-density electroencephalogram EEG of 15 subjects, which is collected using a 62-channel electrode cap with a sampling rate of 1000 Hz. The task is to identify three emotional states (positive, neutral, and negative). The original high-density electroencephalogram EEG is preprocessed by band-pass filtering (0.5-70 Hz), down-sampling (down to 200 Hz), and segmenting (2 seconds per sample), etc. to generate a training set and a test set. Parameter settings: original total channel number C h = 62, target channel number C l = 8, time series length T = 400 (2 seconds x 200 Hz), batch size B = 32, and number of classes K = 3.

[0114] The original high-density electroencephalogram signal contains 32 samples, each sample includes 62 channels, and each channel includes 400 time points (i.e. time series length).

[0115] Step S2: Channel importance score and adaptive selection. The convolution kernel size k t of the one-dimensional convolution = 16, the step size d = 2, and the number of filters Nf = 32; the convolution output is the time feature where the dimension of the time feature The global pooling is the compressed time feature The compressed time dimension; the multi-layer perception (MLP) calculates the importance score of each channel That is, the importance score of 62 channels of each sample; the history memory updates the parameter as a = 0.9, Y = 0.1; the TopK selects the 8 channels with the highest importance score, and the channel index is obtained For example, assuming that the selected 8 channel indexes are [3, 8, 12, 24, 31, 42, 53, 59]: the final output is The 8 key channel data of each sample are retained;

[0116] Step S3: feature extraction and embedding. The convolution kernel size k t = 8, the filter number F = 64, and the output of the time convolution is the local time feature where 393 = 400 - 8 + 1; and the output of the spatial convolution is the fused spatial feature Let the pooling kernel size k p = 4, and the pooling output is the feature dimension where Let the embedding dimension of the projection E = 128, and the final embedding feature is

[0117] Step S4: collaborative guidance learning. Task-aware branch: after the global pooling, the output is The full connection layer is FC3 (128 x 64) and FC4 (64 x 3); the classification output is The probability of each sample belonging to three emotions. Task-agnostic branch: the electroencephalogram signal mapped back is obtained The channel super-resolution graph convolution layer parameter is C l = 8, C h = 62, and thus the reconstructed electroencephalogram signal is obtained The original high-density signal is restored.

[0118] Step S5: joint loss function optimization. The reconstruction loss is (an example value); the classification loss is (an example value); the diversity loss is (an example value); and the weight parameter is l class = 1.0, and l div = 0.1; and the total loss is

[0119] Through the experiment verification on the SEED data set, the method successfully reduces the channel number from 62 to 8 (reducing 87%), and the classification accuracy only slightly decreases from 91.3% to 89.2% (decreasing 2.1%). At the same time, the signal reconstruction quality remains good, with an average correlation coefficient of 0.86, indicating that most of the original signal information is retained. In terms of computing efficiency, compared with full-channel processing, the inference speed is improved by 4.8 times after using the method, and the memory occupancy is reduced by about 75%, proving the significant advantages of the EEG channel selection method in the application.

[0120] The application further provides an EEG channel selection system based on a dual-branch cooperative guidance learning in another aspect, comprising a device end.

[0121] A high-density electroencephalogram signal acquisition unit is configured to receive original high-density electroencephalogram signals EEG from electrodes and pre-process the original high-density electroencephalogram signals EEG.

[0122] A signal processing unit is connected to the high-density electroencephalogram signal acquisition unit and internally stores a trained electroencephalogram channel selection model, and the signal processing unit is configured to realize real-time calculation based on an embedded processor or a field programmable gate array (FPGA).

[0123] A human-computer interaction unit is electrically connected to the signal processing unit and configured to provide a touch screen or button interface for an operator to control and view results.

[0124] The human-computer interaction unit is internally provided in the application, and the human-computer interaction unit enables the application to have channel selection visualization capability, facilitating neuroscientific research and analysis.

[0125] The above description is only a specific embodiment of the application, but the protection scope of the application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the application, which should be covered within the protection scope of the application. Furthermore, the technical solutions of each embodiment of the application can be combined with each other, but it must be based on the realization of a person skilled in the art, and when the combination of technical solutions appears contradictory or unachievable, it should be considered that the combination of technical solutions does not exist and is not within the protection scope of the application. Therefore, the protection scope of the application should be subject to the protection scope of the claims.

Claims

1. A method for EEG channel selection based on dual-branch collaborative guided learning, characterized in that, Includes the following steps: S1. Collect the raw EEG signal and then construct an EEG channel selection model. The EEG channel selection model includes a signal input module, a channel importance assessment module, a signal feature extraction module, a collaborative guidance learning module, and a channel selection output module connected in sequence. S2. The raw EEG signal is input into the channel importance assessment module through the signal input module. The channel importance assessment module calculates the smoothing score of each channel and selects a set number of optimal channels based on the smoothing score of each channel. S3. Input the selected optimal channel into the signal feature extraction module, extract the spatiotemporal characteristics of the optimal channel, and then perform pooling and projection to obtain the final embedded features. S4. Input the embedded features into the task-aware branch and the task-independent branch in the collaborative guidance learning module respectively. In the task-aware branch, perform global average pooling on the embedded features and obtain the predicted label probability distribution through multi-class prediction. In the task-independent branch, reconstruct the complete EEG signal based on the embedded features to obtain the reconstructed EEG signal. S5. Based on the predicted label probability distribution and the reconstructed EEG signal, construct the total loss function, iterate from S2 to S4, minimize the total loss, and adjust the weights of the EEG channel selection model during the iteration process to obtain the trained EEG channel selection model. S6. Deploy the trained EEG channel selection model to the device, and use the trained EEG channel selection model to select channels in the actual EEG signal to obtain the channel selection matrix, thereby completing the selection of EEG channels.

2. The EEG channel selection method based on dual-branch collaborative guided learning according to claim 1, characterized in that, The channel importance evaluation module includes a one-dimensional convolutional neural network Conv1D, a global average pooling layer, and a multilayer perceptron MLP connected in sequence, wherein the multilayer perceptron MLP includes a fully connected layer and a modified linear unit activation function connected in sequence. The signal feature extraction module includes a spatiotemporal convolutional network ConvEEG, a linear unit activation function ELU, a pooling layer, and a projection layer connected in sequence. The collaborative guided learning module includes two parallel branches: a task-aware branch and a task-independent branch. The task-aware branch includes a global average pooling layer and a classifier connected in sequence. The task-independent branch includes an anti-embedding network and a channel super-resolution network connected in sequence. The anti-embedding network includes a transposed convolutional layer and an upsampling layer connected in sequence. The channel super-resolution network includes a graph convolutional network and a feedforward network connected in parallel.

3. The EEG channel selection method based on dual-branch collaborative guided learning according to claim 1, characterized in that, S2 specifically includes the following steps: S21. The original EEG signal is input into the channel importance assessment module through the signal input module. The channel importance assessment module models the information of each channel through a one-dimensional convolutional neural network Conv1D and a multilayer perceptron MLP, calculates the importance score of each channel, and then calculates the smoothing score of each channel based on the importance score of each channel. S22. Based on the smoothed score of each channel, and using the Gumbel-TopK mechanism, select C. l The optimal channel.

4. The EEG channel selection method based on dual-branch collaborative guided learning according to claim 3, characterized in that, S21 specifically includes the following steps: S211. The raw EEG signal is input into the channel importance assessment module via the signal input module. The one-dimensional convolutional neural network Conv1D in the channel importance assessment module extracts features from the time series of each channel to obtain the temporal features, as follows: Z=Conv1D(X h ,k t ,d,N f ) Where Z represents the time feature, and Let B represent the set of real numbers, B represent the batch size, and C represent the set of real numbers. h T represents the total number of channels used in EEG acquisition, i.e., the number of electrodes; T′ represents the third dimension of the time feature, and X h This represents the original electroencephalogram (EEG) signal, and T represents the time series length of each channel; k t N is the temporal kernel size; d is the convolution stride; N f Number of convolutional filters: S212. The dimension of the time feature Z is compressed using a global average pooling layer to obtain the compressed time feature Z′, as follows: Where GlobalAvgPool represents the dimension compression operation of the global average pooling layer; S213. Based on the compressed time feature Z′, and using a multilayer perceptron (MLP), the importance score for each channel is calculated as follows: Where f represents the channel importance score, and FC1 and FC2 are respectively based on dimension C. h ×64 and 64×C h A fully connected layer; ReLU is a modified linear unit activation function; S214. Introduce a historical memory mechanism to smooth the importance score of each channel, resulting in a smoothed score for each channel, as follows: in, The channel's smooth score is represented by γ, which represents the historical impact factor, and h is the channel's smooth score. b-1 This represents the historical channel selection frequency in round b-1; and h b-1 The following relationship must be satisfied: Among them, h b Select the frequency for the historical channel in round b; α is the momentum update coefficient; S b For the channel selection matrix in round b, This indicates the desired operation, which is to average the batch dimension.

5. The EEG channel selection method based on dual-branch collaborative guided learning according to claim 4, characterized in that, S22 specifically includes the following steps: S221. Add Gumbel noise to the smoothed score of each channel and perform a softmax operation to obtain the selection probability p of each channel. b The details are as follows: Among them, g b =-log(-log(u) b )) represents Gumbel noise; u b ~Uniform(0,1) is a random variable sampled from a uniform distribution; Let represent the smoothed score of the b-th batch; τ represents the randomness of the temperature parameter control sampling, τ is adjusted using an exponential annealing strategy, and the formula for calculating τ is: Where, τ init t is the initial temperature; β is the annealing rate; t is the current training step; exp represents an exponential function with the natural constant e as the base. S222, the selection probability p for each channel b Perform a TopK operation to obtain the index of each channel, as follows: I b =TopK(p b ,C l ) in, The index of the selected channel; S223. Select the top C channels with the highest probabilities based on the index of each channel. l Each channel yields the selected EEG signal, as follows: X l [b,i,:]=X h [b,I b [i],:] in, This indicates that only the top C-ranked EEG signals were retained after selection. l Bit-level channel data.

6. The EEG channel selection method based on dual-branch collaborative guided learning according to claim 5, characterized in that, S3 specifically includes the following steps: S31, Select the EEG signal X l The signal is input into the signal feature extraction module, where the spatiotemporal convolutional network ConvEEG processes the selected EEG signal X. l Perform temporal convolution to obtain local temporal features X temp The details are as follows: Where, k t The size of the convolution kernel over time; The learnable temporal convolutional kernel weights, where c represents the channel number and F represents the number of temporal filters; f b This indicates the filter number, and t is the time index; S32. Perform spatial convolution along the channel dimension, and apply it based on the local temporal feature X. temp By fusing spatial information from different channels, the fused spatial feature X is obtained. spatial The details are as follows: in, Spatial convolution kernel; S33. Spatial features X after compression and fusion using the exponential linear unit activation function ELU and average pooling. spatial The pooled features are obtained. Specifically as follows: Where, k p t is the pooling kernel size; t′ is the third-dimensional index of the feature after pooling; S34. The pooled feature X... pooled The final embedded features are obtained by projecting them into the embedding space through 1×1 convolution. Specifically as follows: Where e is the embedding dimension index, Let E be the projection matrix and E be the embedding dimension.

7. The EEG channel selection method based on dual-branch collaborative guided learning according to claim 6, characterized in that, S4 specifically includes the following steps: S41. Embed feature X patch The inputs are respectively fed into the task-aware branch and the task-independent branch of the collaborative guidance learning module. If it is a task-aware branch, then proceed to S42; otherwise, proceed to S43. S42. Within the task-aware branch, firstly, the embedded feature X... patch Perform global average pooling, then input the data into a classifier for multi-class prediction to obtain the predicted label probability distribution. S43. In the task-independent branch, the embedded features X are first embedded through the de-embedding network. patch Mapping back to the time series yields the mapped EEG signals. Based on the mapped EEG signals Reconstructing the complete EEG signal yields the reconstructed EEG signal.

8. The EEG channel selection method based on dual-branch collaborative guided learning according to claim 7, characterized in that, The entire process of S42 is represented by a formula, as follows: Where g represents the pooled embedded features; Dropout represents the dropout layer; softmax represents the normalization exponential function; FC3 and FC4 are fully connected layers with dimensions E×128 and 128×K, respectively; and K is the number of classes. To predict the probability distribution of labels; The mapped EEG signals in S43 The formula is as follows: Among them, the inverse embedding network DeConvEEG is a transposed version of the spatiotemporal convolutional network ConvEEG, which includes transposed convolution, upsampling and other operations; The reconstructed EEG signal in S43 The formula is as follows: SuperResolution is a channel super-resolution network.

9. The EEG channel selection method based on dual-branch collaborative guided learning according to claim 8, characterized in that, The total loss function in S5 is as follows: in, Indicates the total loss; Indicates the losses incurred during reconstruction; Indicates classification loss; Indicates the loss of diversity; λ class , λ div These represent the hyperparameters for classification loss and diversity loss, respectively, and are used to adjust the weights between different loss terms; The reconstruction loss The formula for calculation is: Where ||·||2 is the L2 norm, used to measure the mean square error between the original EEG signal and the reconstructed EEG signal; The classification loss The formula for calculation is: Where yb,k is the unique hot code of the true label of the k-th category in the b-th batch. Let K be the probability of the predicted label for the k-th category in the b-th batch, where K is the number of categories. The loss of diversity The formula for calculation is: Among them, s ：，c Let represent the selection probability of the c-th channel in different samples, and std(·) be the standard deviation.

10. An EEG channel selection system based on dual-branch collaborative guided learning, characterized in that, The device includes a terminal that selects EEG channels using the EEG channel selection method shown in any one of claims 1 to 9. The terminal includes: The EEG signal acquisition unit is responsible for receiving the raw EEG signals from the electrodes and preprocessing the raw EEG signals. The signal processing unit is connected to the EEG signal acquisition unit and has a built-in trained EEG channel selection model. The signal processing unit implements real-time computing functions based on an embedded processor or a programmable gate array (FPGA). The human-machine interface unit is electrically connected to the signal processing unit and is used to provide a touch screen or button interface for the operator to control and view the results.

Images