An electroencephalogram signal recognition method based on an information separator and a region convolution network

Abstract

The application claims a kind of electroencephalogram signal recognition method based on information separator and regional convolution network, comprising: collecting electroencephalogram data, and carrying out filtering, wavelet packet decomposition and data standardization and other pretreatment to electroencephalogram data;According to the electrode installation position, a topological graph is constructed, and then the preprocessed electroencephalogram data is embedded in the graph;Using the constructed graph to create an information separator to obtain independent source information for each node, then using a regional convolution network that focuses on regional features and sequentially increases the range of the region of interest to extract electroencephalogram signal features, the regional convolution network includes three convolution layers, which are node convolution layer for extracting node features, domain convolution layer for extracting domain features centered on the node, and global convolution layer for extracting global features;Finally, classification and recognition are completed by a fully connected layer.The application can significantly reduce the information redundancy between electrodes and effectively improve the recognition accuracy of electroencephalogram signals.

Description

Technical Field

[0001] This invention relates to the field of smart biomedicine, and in particular to a method for recognizing electroencephalogram (EEG) signals based on an information separator and a regional convolutional network. Background Technology

[0002] Brain-computer interface (BCI) technology is a human-computer interaction technology that enables direct communication between the human brain and computers or other electronic devices. The human brain contains tens of billions of neurons, and electroencephalogram (EEG) signals are the electrical signals generated by the activity between these neurons. EEG signals have been widely used in fields such as artificial limbs, sports rehabilitation, brain disease diagnosis, and fatigue detection. When the brain engages in thought activities, it generates specific electrical signals that are mapped to the cerebral cortex. Decoding human intentions based on these electrical signals is the key to brain-computer interfaces.

[0003] When people are physically moving or imagining moving, event-related desynchronization (ERD) and event-related synchronization (ERS) phenomena can be observed, which help decode human intentions. Therefore, motor imagery is currently a hot research topic in the field of EEG signals, and increasing research is making the decoding of motor imagery EEG signals more reliable. EEG signals possess most of the characteristics of bioelectrical signals, such as nonlinearity, instability, and randomness, and also contain artifacts and noise from other bioelectrical signals. The amplitude of this noise is several times or even tens of times greater than that of the EEG signal. Therefore, traditional methods cannot extract effective features from EEG signals. In recent years, due to the rapid development of deep learning and its powerful processing capabilities for random and nonlinear data, it has gradually become one of the mainstream methods in EEG research.

[0004] Currently, deep learning-based methods for EEG signal intention decoding primarily treat the raw EEG signal as a two-dimensional matrix, with the two dimensions being the electrodes and the acquired EEG time series, respectively. Deep learning algorithms, such as Convolutional Neural Networks (CNNs) or Long Short-Term Memory (LSTMs), are then used for feature extraction and classification. These methods have achieved relatively good performance. However, it's worth noting that these methods use the raw EEG data represented by a two-dimensional matrix as input data. However, given that the brain can be considered a sphere with weak electrical signals, and the cerebral cortex is a continuous potential surface, the electrical signal from each electrode will contain information from the locations of other electrodes. Therefore, directly using the raw data leads to information redundancy. Independent Component Analysis (ICA), as an important tool for blind source separation, can effectively reduce this information redundancy. However, for EEG data, the number of sources is uncertain, and ICA can disrupt the original electrode order, causing the data to lose its original spatial information. Therefore, constructing a graph can effectively avoid this problem.

[0005] Graph Convolutional Networks (GCNs) are a classic method for graph data processing. They aggregate neighbor information at each layer using a symmetric normalized adjacency matrix, and then perform feature transformation using a weight matrix shared by each node. They have achieved good results in semi-supervised node classification and some graph classification tasks. Simplified Graph Convolutional Neural Networks (SGCNs) are an improvement on GCNs. Compared to GCNs, they remove the feature transformation at each layer but achieve almost the same results. This illustrates that the aggregation of neighbor information plays a major role in GCNs. These two methods provide a framework for graph data processing: utilizing the adjacency relationships between nodes to generate new node information. On the other hand, when the brain engages in thought activities, the activity levels of neurons in different brain regions vary greatly. However, current methods mostly focus on the overall spatial information of all neurons, neglecting the need to pay more attention to the spatial information of each local region in practical applications.

[0006] CN113128552A discloses a method for EEG emotion recognition based on a deep separable causal graph convolutional network. Existing neural networks for identifying EEG emotions lack consideration of the directional functional relationships between channels. This invention proceeds as follows: 1. Collecting EEG data from the subject. 2. Constructing an adjacency matrix. 3. Calculating the regularized Laplacian matrix corresponding to the EEG data. 4. Using a deep separable causal graph convolutional model for feature extraction and emotion classification. This invention employs Granger causality to model the relationships between EEG signal channels, constructing a directed asymmetric matrix that fully considers the causal relationships between channels, consistent with the actual generation of EEG signals, thus effectively improving the accuracy of emotion type recognition. Furthermore, this invention uses deep separable convolution to fully extract local features from the EEG data, further improving classification accuracy.

[0007] This patent uses Chebyshev polynomials based on regularized Laplacian matrices for graph convolution operations to achieve EEG emotion recognition. However, this patent treats directly collected raw data as EEG features, ignoring the high information redundancy between electrodes due to the conductivity of electrical signals, thus failing to effectively extract EEG data features. Furthermore, graph convolution operations based on Chebyshev polynomials only generate new node features within the node's neighborhood by repeatedly aggregating neighbor information, lacking attention to features of different spatial ranges, and the feature extraction capability of graph convolution operations based on Chebyshev polynomials is also relatively limited. This patent, on the other hand, uses an information separator to obtain independent source information of nodes to reduce information redundancy between nodes, and then uses a region convolutional network with the spatial range of interest increasing sequentially from node to neighborhood to global to fully extract features of different spatial ranges. Summary of the Invention

[0008] This invention aims to solve the problem of low EEG signal recognition rate in existing technologies. It proposes an EEG signal recognition method based on an information separator and a region convolutional network. The technical solution of this invention is as follows:

[0009] A method for EEG signal recognition based on an information separator and a region convolutional network includes the following steps:

[0010] Step 1: Collect EEG data and perform preprocessing on the EEG data, including filtering, wavelet packet decomposition, and data standardization;

[0011] Step 2: Construct a topology map based on the electrode installation locations, and then embed the preprocessed EEG data into the topology map;

[0012] Step 3: Based on the constructed topology graph, create a transformation matrix that can obtain the source information of the nodes and call it the information separator. Then use the information separator to obtain the independent source information of each node.

[0013] Step 4: Extract the node's own features using the node convolutional layers of the region convolutional network;

[0014] Step 5: Extract neighborhood features centered on nodes using the neighborhood convolutional layers of a region convolutional network;

[0015] Step 6: Use graph pooling to reduce the graph size;

[0016] Step 7: Use the global convolutional layer of the region convolutional network to extract global features of all functional region blocks and complete the classification;

[0017] Step 8: Set the loss function: cross-entropy loss function;

[0018] Furthermore, the preprocessing of the EEG data, including filtering, wavelet packet decomposition, and data standardization, specifically includes:

[0019] The EEG data was bandpass filtered from 0.5 to 100 Hz. First-order wavelet packets were used to decompose the filtered data into approximate and detail components. The approximate component is low-frequency, and the detail component is high-frequency. Based on the rhythmic characteristics of EEG, the detail component was discarded as noise, and only the approximate component was retained. Z-score normalization was used to standardize the decomposed approximate data. The Z-score normalization formula is as follows: Where x represents the approximate data from the decomposition, and μ is the sample mean of the approximate data. It is the standard deviation of the data, N is the sample size, and x is the standard deviation of the data. z It is standardized data, x z The mean is 0 and the standard deviation is 1.

[0020] Furthermore, the step of constructing the diagram based on the electrode mounting positions includes:

[0021] Each electrode is treated as a node, and the preprocessed electrode data is used as the node feature. All nodes form a node set N. A distance threshold is set, and nodes whose distance to the target node is less than the threshold are selected as the neighbor of that node. Since the closer the distance in the electric field, the greater the connection, the weight of the edge between adjacent nodes is set to the reciprocal of the distance. All edges form an edge set V. The constructed feature topology graph is: G = {N, V}.

[0022] Furthermore, the step of creating an information separator based on the constructed topology graph to obtain independent source information for each node specifically includes:

[0023] Based on the constructed graph, the elements of the adjacency matrix A of this graph are:

[0024]

[0025] Where, d ij Let be the distance between node i and node j. Then the Laplacian matrix of the graph is: L = DA, where Let L be the degree matrix of A; the divergence of the gradient of a scalar field represents the intensity of the sources in the field. The divergence of the gradient of a scalar field is represented by the Laplace operator, and the Laplace matrix is ​​the discrete form of the Laplace operator. Therefore, the Laplace matrix is ​​used to obtain the node-independent source information; to preserve the magnitude of the data, the Laplace matrix is ​​standardized using the random walk Laplace matrix: L rw =L / D=IA / D, L rw The elements are:

[0026]

[0027] Add trainable parameters to fully capture node source information: Where W is a weight matrix with the same shape as A, I is the identity matrix, and "·" denotes the dot product operation, then the source information of the node can be represented as:

[0028] Furthermore, the node convolutional layer steps of the region convolutional network include:

[0029] First, information from different time periods at each node is extracted using one-dimensional convolution. The formula for one-dimensional convolution is: in, S represents the j-th feature of node c in layer 1. c,j It is on node c The input feature set, f(·) and f(·) are the connection weights, biases, and ELU activation functions, respectively; the topological structure of the feature topology map remains unchanged after convolution, and max pooling is used to reduce the oversensitivity of the convolutional layer to the temporal location.

[0030] Furthermore, the neighborhood convolutional layer step of the region convolutional network includes:

[0031] First, a common convolutional kernel with a graph structure is established, whose topology is the same as the feature topology graph. Then, this convolutional kernel is used to convolve the features in the graph output by the convolutional layer around local regions centered on each node. The specific operation is as follows:

[0032] For each node, select the features of the node itself and its neighbors, along with the convolutional kernel channels, from the common convolutional kernel. Reassemble the selected features and convolutional kernel channels into a matrix in the same order. For the reassembled features, considering that different regions have different levels of attention to different nodes, node attention is used to obtain the attention coefficients of the current region block for different member nodes. The steps for calculating the node attention coefficients are as follows: ① Calculate the average feature value of the node: Where L is the number of features of a node, u i Let W1 be the i-th feature of the node; ② Calculate the attention coefficients of the current region block for different nodes as: S = σ(W2δ(W1u)), where W1 and W2 are the weights of the linear transformation matrix, and σ and δ are the sigmoid and ELU activation functions, respectively; thus, the features of the current region block after reorganization are:

[0033] Then, the recombined convolutional kernel is used to perform a convolution operation on the recombined features. The operation formula for this convolutional layer is: Where, N c It is a region block centered on node c, that is, node c and its neighbors. S represents the j-th feature of node c in the second layer. m,j It is on node c The input feature set, f(·) represents the connection weights, biases, and ELU activation function, respectively; finally, the regional features of each node are re-embedded into the graph.

[0034] The graph pooling technique is used to pool the graph after convolution and extract features from the global region.

[0035] Furthermore, the graph pooling steps are as follows: ① Select the node closest to the geometric center among the 8 functional regions as the center node, and all nodes in the functional regions as member nodes; ② Calculate the distance between the member nodes and the center node: Where n is the number of nodes contained in the current functional area block (including the central node), a iIt is the i-th member node of the current region block. and These are the j-th node of the central node and the j-th feature of the i-th subordinate node, respectively; ③ Calculate the new node representing the current functional area block, and the features of the new node are:

[0036]

[0037] Among them, h i The feature of the i-th node in the current region block. It represents the contribution of the i-th node to the new node.

[0038] Furthermore, the global convolutional layer step of the region convolutional network includes:

[0039] A convolutional kernel with the same structure as the pooled graph is established to extract global region features from eight functional regions. The operation formula for this convolutional layer is as follows: Where N represents the set of all functional area blocks, S represents the j-th feature of the 3rd layer. m,j It is on node c The input feature set, f(·) represents the connection weights, bias, and ELU activation function, respectively; finally, a fully connected layer and softmax activation function are used to obtain the final classification result.

[0040] Furthermore, the loss function is the cross-entropy loss function: Where c is the number of categories, y is the true label, and p is the predicted probability of the category.

[0041] The advantages and beneficial effects of this invention are as follows:

[0042] This invention constructs EEG signals as a graph, and then obtains independent source information of nodes from the Laplacian matrix of the graph, which greatly reduces information redundancy and improves information utilization. At the same time, it uses a regional convolutional network to fully extract features from different regions from nodes to local to global, which significantly improves the recognition accuracy of EEG signals.

[0043] The innovation of this invention lies primarily in step 3, which uses a random walk Laplace matrix combined with a weight matrix constrained by the adjacency matrix to generate a transformation matrix capable of acquiring node source information. The difference between the random walk Laplace matrix and the symmetric standardized Laplace matrix is ​​that the random walk Laplace matrix discards the symmetry characteristic and uses the node's own degree for normalization, making it more conducive to extracting node source information. The weight matrix constrained by the adjacency matrix can avoid interference from some difficult-to-measure factors, such as differences in electrode connection quality. Furthermore, step 4, considering the uncertainty and periodicity of the specific start time of motor imagery, uses node convolutional layers to fully extract the node's own features. Step 5 establishes a common convolutional kernel with the same topological structure as the feature topology graph, selects and reorganizes the features and convolutional kernels of the region blocks centered on each node to complete the convolution operation, and considers the different attention levels of each region block to different nodes during feature reorganization, using node attention to calculate the attention coefficients of different region blocks to different nodes. Step 6 uses graph pooling technology to reduce the graph size based on the division of functional areas of the cerebral cortex. Step 7 uses a global convolutional layer to extract global information from all functional regions for final classification, thus completing the recognition of EEG signals. Attached Figure Description

[0044] Figure 1 This is a flowchart of a preferred embodiment of the EEG recognition method based on an information separator and a regional convolutional network provided by the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.

[0046] The technical solution of the present invention to solve the above-mentioned technical problems is:

[0047] Please see Figure 1 The present invention provides an EEG recognition method based on an information separator and a regional convolutional network, comprising the following steps.

[0048] Step S01: EEG Data Acquisition. EEG signal data acquisition was performed using a 32-channel BRAINPRODUCTS EEG acquisition system. Electrodes were placed according to the international standard electrode method, with a sampling frequency of 128Hz. The acquired EEG signals were transmitted to a computer via wired connection. The experiment began at t=0, with a buzzer prompting for 0.3 seconds and a fixed crosshair appearing on the screen for 2 seconds to prompt the subject to concentrate. Starting at t=2s, the motor imagery task was prompted, and the subject began motor imagery for 4 seconds, followed by a 4-second rest period after the screen went black. The motor imagery task consisted of four parts: movements of the left hand, right hand, tongue, and feet. Data preprocessing involved bandpass filtering the EEG data from 0.5-100Hz. First-order wavelet packets were used to decompose the filtered data into approximate (low-frequency) and detailed (high-frequency) components. Based on the rhythmic characteristics of EEG, the detailed components were discarded as noise, retaining only the approximate components. Since EEG signals are bipolar data and the maximum and minimum values ​​are difficult to estimate, the Z-Score standardization method is used to standardize the approximate data after decomposition. The Z-Score standardization formula is: Where x represents the approximate data from the decomposition, and μ is the sample mean of the approximate data. It is the standard deviation of the data, N is the sample size, and x is the standard deviation of the data. z This is standardized data. z The mean is 0 and the standard deviation is 1.

[0049] Step S02, Graph Construction. The steps for constructing a feature topology graph based on electrode installation locations include: treating each electrode as a node, using preprocessed electrode data as node features, resulting in a node set N. A distance threshold is set, and nodes whose distance to the target node is less than the threshold are selected as its neighbors. Since closer proximity in an electric field increases connectivity, the weight of the edges between adjacent nodes is set to the reciprocal of the distance. All edges form an edge set V. The constructed graph is: G = {N, V}.

[0050] Step S03: Obtain node source information. The steps for creating an information separator based on the graph and obtaining node source information include:

[0051] Based on the constructed graph, the elements of the adjacency matrix A of this graph are:

[0052]

[0053] Where, d ij Let be the distance between node i and node j. Then the Laplacian matrix of the graph is: L = DA, where This is the degree matrix of A. The divergence of the gradient of a scalar field represents the intensity of the sources in the field. The Laplace operator is calculated as the divergence of the gradient. The Laplace matrix is ​​the discrete form of the Laplace operator, so the Laplace matrix can be used to obtain node-independent source information. To preserve the magnitude of the data, the Laplace matrix is ​​standardized using a random walk Laplace matrix: L rw =L / D=IA / D, L rw The elements are:

[0054]

[0055] Considering some difficult-to-measure factors, such as the varying connection quality of each electrode, trainable parameters are added to fully capture node source information: Where W is a weight matrix with the same shape as A, I is the identity matrix, and "·" denotes the dot product operation. The source information of a node can then be represented as... Step S04: Extract temporal features from nodes. The steps for extracting temporal features from nodes in the node convolutional layer of the region convolutional layer include:

[0056] Because the mental activity of subjects during motor imagery is periodic and the precise start time of motor imagery is difficult to determine, we first extract information from different time periods at each node using one-dimensional convolution. The formula for this convolution is as follows: in, This represents the j-th feature of node c in layer 1. c,j It is on node c The input feature set, f(·) and f(·) are the connection weights, biases, and ELU activation functions, respectively. The topological structure of the feature topology map remains unchanged after convolution, and max pooling is used to reduce the oversensitivity of the convolutional layer to temporal location.

[0057] Step S05: Extract local region features. The steps for extracting local features from the local convolutional layer of the region convolutional layer include:

[0058] First, a common convolutional kernel with a graph structure is established, whose topology is the same as that of the feature topology graph. Then, this convolutional kernel is used to convolve the features in the graph output by the convolutional layer in claim 5 around the local regions centered on each node. The specific operation is as follows:

[0059] For each node, select the features of the node itself and its neighbors, along with the convolutional kernel channels, from the common convolutional kernel. Reassemble the selected features and convolutional kernel channels into a matrix in the same order. For the reassembled features, considering that different regions have different levels of attention to different nodes, node attention is used to obtain the attention coefficients of the current region block for different nodes. The steps for calculating the node attention coefficients are as follows: ① Calculate the average feature value of the node: Where L is the number of features of a node, u i Let be the i-th feature of the node. ② Calculate the attention coefficients of the current region block for different nodes as: S = σ(W2δ(W1u)), where W1 and W2 are the weights of the linear transformation matrix, and σ and δ are the Sigmoid and ELU activation functions, respectively. Therefore, the features of the current region block after reorganization are:

[0060] Then, the recombined convolutional kernel is used to perform a convolution operation on the recombined features. The operation formula for this convolutional layer is: Where, N c It is a region block centered on node c, that is, node c and its neighbors. S represents the j-th feature of node c in the second layer. m,j It is on node c The input feature set, f(·) represents the connection weights, biases, and ELU activation function, respectively. Finally, the regional features of each node are re-embedded into the graph.

[0061] The human cerebral cortex can be divided into four main functional areas (frontal, parietal, temporal, and occipital lobes). Based on the left and right hemispheres, the cerebral cortex can be further divided into eight functional regions. Graph pooling can then be used to pool the convolutional graph into smaller graphs to facilitate the computation and extraction of global region features. The graph pooling steps are as follows: ① Select the node closest to the geometric center among the eight functional regions as the center node, and the other nodes in the functional regions as auxiliary nodes. ② Calculate the distance between the auxiliary nodes and the center node. Where n is the number of nodes contained in the current functional area block, a i This refers to the i-th node (inclusive) of the current region block. and These are the j-th node of the central node and the j-th feature of the i-th node, respectively. ③ Calculate the new node representing the current functional area block. The features of the new node are:

[0062]

[0063] Among them, h i The feature of the i-th node in the current region block. It represents the contribution of the i-th node to the new node.

[0064] Step S06: Extract global features from the global region and classify them. The steps for extracting global features from the global convolutional layer of the region convolutional layer include:

[0065] A convolutional kernel with the same structure as the pooled graph is established to extract global region features from eight functional regions. The operation formula for this convolutional layer is as follows: Where N represents the set of all functional area blocks, S represents the j-th feature of the 3rd layer. m,j It is on node c The input feature set, f(·) represents the connection weights, biases, and ELU activation function, respectively. Finally, a fully connected layer and the Softmax activation function are used to obtain the final classification result.

[0066] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions.

[0067] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0068] The above embodiments should be understood as illustrative only and not as limiting the scope of protection of the present invention. After reading the description of the present invention, those skilled in the art can make various alterations or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A method for electroencephalogram signal recognition based on information separator and region convolution network, characterized in that, The method comprises the following steps: Collecting electroencephalogram data and performing preprocessing on the electroencephalogram data, including filtering, wavelet packet decomposition and data standardization; Constructing a topology graph according to the electrode installation position, and embedding the preprocessed electroencephalogram data into the topology graph; According to the constructed topology graph, a conversion matrix capable of obtaining node source information is created and is called an information separator, and the information separator is used to obtain independent source information of each node; Then, a region convolutional network with focused region features and an area range of interest that increases sequentially is used to extract electroencephalogram signal features, and the region convolutional network comprises three convolutional layers, namely a node convolutional layer for extracting node features, a domain convolutional layer for extracting domain features centered on the node, and a global convolutional layer for extracting global features; Finally, a full connection layer is used to complete classification and recognition; The step of constructing a feature topology graph according to the electrode installation position comprises: Taking each electrode as a node, taking the pre-processed electrode collection data as the node feature, all nodes constitute a node set ; setting a distance threshold, selecting the nodes with a distance less than the threshold to the target node as the neighbors of the target node, since in the electric field, the closer the distance, the greater the contact, so the weight of the edge of the adjacent node is set as the inverse of the distance, all edges constitute an edge set , the constructed topology graph is: ; The step of obtaining independent source information of each node by using the information separator created according to the constructed topology graph comprises: From the constructed topology graph, the adjacency matrix of this graph is derived The elements of the matrix are: wherein, is a node and a distance between the node and the node , a Laplacian matrix of the topological graph is: wherein, is a degree matrix of the node ; a divergence of a gradient of a scalar field represents an intensity of a source in the field, the divergence of the gradient of the scalar field is represented by a Laplacian operator, the Laplacian matrix is a discrete form of the Laplacian operator, and thus the source information independent of the node is obtained by using the Laplacian matrix; in order to maintain a magnitude of data, a random walk Laplacian matrix is used to perform a standardization processing on the Laplacian matrix: , an element of the random walk Laplacian matrix is: Add trainable parameters to fully capture the node source information: where, is a weight matrix of the same shape as , is an identity matrix, denotes the dot product operation; and is named as information separator; then the source information of a node can be represented as .

2. The electroencephalogram signal recognition method based on information separator and region convolution network according to claim 1, characterized in that, The preprocessing of the electroencephalogram data, including filtering, wavelet packet decomposition and data standardization, comprises: The EEG data is band-pass filtered at 0.5-100 Hz, and the filtered data is decomposed into an approximation part and a detail part using a first-order wavelet packet, the approximation part is low frequency, and the detail part is high frequency, according to the rhythm characteristics of the EEG, the detail part is discarded as noise, and only the approximation part is retained; the decomposed approximation part data is standardized using a Z-Score standardization method, and the Z-Score standardization formula is: wherein, is the decomposed approximation part data, is the sample mean of the approximation part data, is the standard deviation of the data, is the sample number, is the standardized data, the mean of the standardized data is 0, and the standard deviation is 1.

3. The electroencephalogram signal recognition method based on information separator and region convolution network according to claim 1, characterized in that, The node convolutional layer of the region convolutional network comprises: First, the information of different time periods on each node is extracted by one-dimensional convolution, and the operation formula of one-dimensional convolution is: wherein, represents the node In the first layer of the first characteristic, is the input feature set of the node , and are connection weights, bias and ELU activation function respectively; the topological structure of the feature topology graph after convolution is unchanged and the maximum pooling is used to weaken the excessive sensitivity of the convolution layer to the time position.​​ 4. The electroencephalogram signal recognition method based on information separator and region convolution network according to claim 1, characterized in that, The domain convolutional layer of the region convolutional network comprises: First, a common convolution kernel with a graph structure is established, the topology structure of the common convolution kernel is the same as that of the feature topology graph, then the local region centered on each node in the features in the graph output by the convolution layer is convolved using the common convolution kernel, and the specific operation is as follows: For each node, the features of the node itself and its neighbors and the convolution kernel channels are selected from the public convolution kernel; the selected features and convolution kernel channels are reorganized into a matrix in the same order; for the reorganized features, considering that different regions have different attention degrees to different nodes, the node attention is used to obtain the attention coefficients of the current region block to different nodes, and the coefficient calculation steps of the node attention are as follows: The average value of the features of the node is calculated as follows: wherein, is the number of features of the node, is the feature of the node; The attention coefficient of the current region block to different nodes is calculated as follows: wherein, and are the weights of the linear transformation matrix, and are the sigmod and ELU activation functions respectively; thus, the reorganized features of the current region block are as follows: ; Then the reorganized features are convolved using the reorganized convolution kernel, and the operation formula of the convolution layer is: wherein, is the region block centered on the node , i.e., the neighbor of the node and the node , represents the node at the 2nd layer and the th feature, is the input feature set on the node , , , and are the connection weight, the bias and the ELU activation function respectively; finally, the region features of each node are re-embedded into the graph output by the convolution layer. The graph after convolution is pooled by a graph pooling technology, and the features of the global region are extracted.

5. The electroencephalogram signal recognition method based on information separator and region convolution network according to claim 4, characterized in that, The step of the graph pooling is: Selecting a node closest to the geometric center in the 8 functional region blocks as a center node, and all nodes in the functional region block as membership nodes; Calculate the distance between the membership nodes and the center node: Wherein, is the number of nodes contained in the current functional region block, is the i th membership node of the current region block, and and are the i th node and the i th membership node of the center node respectively, and the i th feature of the i th node and the i th membership node of the center node respectively. ; Calculate a new node representing the current functional region block, and the features of the new node are:​​ wherein, is a feature of a current region block, is a contribution of the th node to the new node.​ 6. The electroencephalogram signal recognition method based on information separator and region convolution network according to claim 5, characterized in that, The global convolutional layer of the region convolutional network comprises: A convolutional kernel with the same structure as the pooled graph is established to extract global region features from eight functional regions. The operation formula for this convolutional layer is as follows: ,in, This represents the set of all functional area blocks. Indicates the third layer One characteristic, It is a node superior The input feature set, , and These are connection weights, biases, and the ELU activation function; finally, a fully connected layer and the softmax activation function are used to obtain the final classification result.

7. The electroencephalogram signal recognition method based on information separator and region convolution network according to claim 6, characterized in that, The loss function is a cross-entropy loss function: wherein is the number of classes, is the true label, is the predicted probability corresponding to the class.

Images