A method for extracting an electroencephalogram signal core network

Abstract

The application provides a brain electrical signal core network extraction method. Through analysis and feature extraction on brain electrical signals, a core node of a brain network is obtained through attention graph convolutional neural network (A-GCN) training, and a core network is also obtained, so that a plurality of classification prediction models only need to calculate a small amount of node or edge features (network) on a test set to obtain higher recognition accuracy. The method provided by the application only needs to collect a small amount of channel data of a subject, and the practicability in BCI application is improved.

Description

Technical Field

[0001] This invention belongs to the fields of electroencephalogram (EEG) signal processing and human-computer interaction, and specifically relates to a method for extracting the core network of EEG signals. Background Technology

[0002] With the advancement of brain science research, the application of brain-computer interfaces (BCIs) has received increasing attention. BCIs combine hardware and software, using voltage fluctuations generated by the human brain to control external devices, such as brain-controlled robotic arms and drones. BCI technology involves multiple disciplines, including neuroscience, human-computer interaction, information processing, and pattern recognition. By preprocessing and extracting features from physiological signals collected from the human brain, it interprets the subject's actual thoughts and converts these thoughts into corresponding machine commands, thereby achieving interactive control between the human brain and the external environment.

[0003] The most effective method for acquiring human brain physiological signals is usually non-invasive electroencephalography (EEG) of the scalp, which has many advantages such as low cost, portability, high temporal resolution, and objectivity and reliability. Although significant scientific achievements have been made in EEG-based task recognition, some challenges remain in practical BCI applications. Current research largely focuses on improving model classification accuracy, neglecting computational complexity, time costs, and the exploration and discovery of key task-related brain regions. The speed of real-time machine feedback significantly impacts user experience; for example, patients with limb movement disorders desperately need timely real-time action feedback. Furthermore, research on core task-related brain regions is becoming increasingly important for a deeper understanding and solution of these problems in the future.

[0004] Complex networks are widely used in time series analysis for feature extraction from various electroencephalogram (EEG) signals. The human brain is a complex network, and brain networks constructed from EEG signals can be used to describe brain activity. The core network contains the main functions of the brain network, and extracting it helps to better understand the complex systems associated with the brain network. Specifically, the core network can be used to analyze the utility of specific brain regions and to remove redundant and interfering information for feature dimensionality reduction, playing a role in the construction and analysis of brain networks. Currently, some manual methods have been used in the field of EEG for core network extraction, which are basically used to calculate the distance between samples using manually defined formulas for feature selection. However, these methods generally lack theoretical foundation, are relatively crude, simplistic, and inflexible, and are not suitable for processing complex samples or multiple sample types.

[0005] In practical applications of BCI, it is necessary to consider not only the specific characteristics of complex samples under a particular experiment, but also the differences between different samples under different experiments. Therefore, developing an efficient method for automatically finding the core network by leveraging the adaptive capabilities of deep learning is of great practical significance. Summary of the Invention

[0006] The first objective of this invention is to overcome the shortcomings of existing technologies and methods, and to provide an efficient and practical automatic core network extraction method. This invention analyzes and extracts features from electroencephalogram (EEG) signals, and uses an attention map convolutional neural network (A-GCN) to train and obtain the core nodes of the brain network. Simultaneously, the core network is also obtained, enabling various classification prediction models to achieve higher recognition accuracy on the test set with only a small number of node or edge features (network). The method provided by this invention requires only the collection of a small amount of channel data from the subject, improving its practicality in BCI applications.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] An automatic method for extracting the core network of EEG signals includes the following steps:

[0009] Step (1): Preprocess the EEG signals, label them with task categories, and construct a dataset:

[0010] 1-1 Data Format Standardization

[0011] Based on the electrode positions of the collected EEG signals, the feature vectors of the total channels are recorded in 1D matrix format in order from left to right and from top to bottom, where N is the total number of channels. At the same time, the EEG signal of each channel is divided into 5 frequency bands to obtain EEG data of 2D tensor data with shape N*5.

[0012] The five frequency bands mentioned are delta (1-4Hz), theta (4-8Hz), alpha (8-14Hz), beta (14-31Hz), and gamma (31-50Hz).

[0013] 1-2 Data Splitting and Cleaning

[0014] The EEG signal obtained in step 1-1 is sliced, and a non-overlapping sliding window operation is performed with T as the time window to obtain a series of EEG signal segments with shape L*N*5, where L is the data length, L=T*w, and w is the sampling frequency.

[0015] Differential entropy (DE) and Spearman rank correlation coefficient (SCC) were calculated for the sliced ​​EEG signal segments. DE was used as node signals and SCC as edge signals (i.e., the correlation between node channels), resulting in a series of N*5 node signals and N*N*5 edge signals. Simultaneously, the sliced ​​EEG signal segments were labeled with their corresponding task categories, ultimately constructing a classification dataset.

[0016] Note that this node signal will not be used as input to the A-GCN model below, and the side signal data for each frequency band will be separated.

[0017] Step (2): Construct an attention-graph convolutional neural network (A-GCN) to extract the core of the EEG network.

[0018] The Attention Graph Convolutional Neural Network (A-GCN) differs from the existing Graph Attention Network (GAT). It is not used to directly train and test classification prediction models, but is a model specifically designed to find core nodes and core networks.

[0019] The A-GCN model takes preprocessed EEG signals of a certain frequency band as input and predicts task category labels as output.

[0020] The A-GCN model includes a cascaded graph signal generation layer, a GCN graph convolutional layer, an attention layer, a flatten layer, a first fully connected layer, and a second fully connected layer.

[0021] To ensure that all information originates solely from the brain network and avoids interference from the node signals within the channels themselves, the graph signal generation layer performs a specific process on each channel (n0, n1, ..., n...). N-1 This generates a unified node signal X(x0, x1, ..., x...). N-1 The node signal X and the input edge signal W∈R N*N Together they form the feature vector of the graph signal;

[0022] The GCN graph convolutional layer receives the graph signal feature vector output by the graph signal generation layer, and performs Fourier transform and Laplacian matrix eigenvalue decomposition on it before convolution calculation to obtain the feature information y after aggregating the information of each neighboring node and the connected edge. Specifically:

[0023] The degree matrix D∈R is calculated from the edge signal W. N*N And obtain the Laplace matrix L = DW ∈ R N*N The characteristic decomposition form of L is UΛU T , where U=[u0,u1,....,u N-1 ]∈R N*NThe columns form a Fourier basis, Λ=diag([λ0,λ1,...,λ N-1 ]) is a diagonal matrix;

[0024] g θ The convolution function is y = g θ (L)x=g θ (UΛU T )x=U g θ (Λ)U T After convolution operation on x, we get y(y0, y1, ... y). N-1 ), y∈R N*1 ;

[0025] The Attention layer receives the output of the GCN graph convolutional layer and multiplies it with an N*1 convolutional kernel to obtain the features processed by the attention mechanism; specifically:

[0026] Feature information y(y0, y1, ... y N-1 The convolution kernel A(a0, a1, ..., a) corresponding to the attention node N-1 Multiplication:

[0027]

[0028] Where A∈R N*1 , This represents element-wise multiplication;

[0029] The Flatten layer receives the features output by the attention layer, which are activated by the ReLU function y = ReLU(y) and flatten the data. Then, the processed data is fully mapped through the first fully connected layer and the second fully connected layer to output the result, and the Softmax function is used to predict the label classification.

[0030] Step (3): Train the A-GCN model in multiple rounds of iterations, and obtain the core node C based on the Attention layer of the A-GCN model after training;

[0031] The loss is calculated after y is compared with the actual labels of the data:

[0032]

[0033] Subsequently, tensor backpropagation is used for gradient descent to adjust the convolution kernel parameters. After multiple rounds of training iterations, the convolution kernel parameters A(a0, a1, ... a2) of the Attention layer are adjusted. N-1 It tends to stabilize.

[0034] After multiple rounds of iterative training of the A-GCN model, the final attention points in the attention layer are sorted in descending order according to the absolute values ​​of the elements in A. This is then applied to the original channels, such as (n5, n0, n3, ...), and the first K channels are selected as the core nodes C(c0, c1, c2, ..., c...). K-1 );

[0035] Step (4): Based on the core node C, retain the corresponding interconnected edges w. ij Where i, j∈(c0, c1, ... c K-1 Finally, the core network S∈R is obtained. K*K .

[0036] Preferably, the graph signal generation layer generates a node signal for each node, with a vector dimension and value of 1, which, together with the input Spearman rank correlation coefficient (SCC) edge signal, constitutes a graph signal.

[0037] Preferably, the kernel size of the GCN graph convolutional layer is 62*62 and 3*1, and the graph convolution order is 3.

[0038] Preferably, the attention convolution kernels of the attention layer correspond to each node at a fixed position, and there are 62 1*1 convolution kernels with an initial value of 1.

[0039] Preferably, the Flatten layer in step (2) will unidimensionalize the multidimensional features activated by the ReLU function.

[0040] Preferably, the convolution kernel parameters of the first fully connected layer and the second fully connected layer in step (2) are 10 convolution kernels of 62*1 and 3 convolution kernels of 10*1, respectively.

[0041] As a preferred option, in step (2), the A-GCN model is run in the PyTorch environment, using the Adam optimizer, with the learning rate lr set to 0.003 and the weight decay weight_decay set to 0.001.

[0042] Generally, traditional brain network analysis employs graph theory metrics such as centrality, shortest path length, clustering coefficient, and local efficiency, focusing on the analysis of complex network topology. Common feature engineering techniques, such as filtering, wrapping, and ensemble methods, offer insights for simplifying brain networks and extracting important information. However, these methods primarily arrange features in rows for comparison and selection using standard one-dimensional feature methods, without utilizing the unique information of neighborhood correlations between channels. Spectrograph theory, on the other hand, has been successfully applied to construct extended graphs, spectral clustering, and graph visualization. Spectrograph filtering, also known as graph convolution, is a popular signal processing method for graph data manipulation, with the Graph Fourier Transform (GFT) being a typical example. Therefore, to extract network information more fully and deeply while ensuring that task-related information comes only from the network, the graph signal generation layer generates a node signal feature vector with a value and dimension of 1 for each node. These node signal vectors, together with the input edge signal vectors, constitute the graph signal feature vector, which is then input into the GCN graph convolutional layer. The graph convolutional layer performs Fourier transform and Laplacian matrix eigenvalue decomposition on this feature vector and then performs convolution calculation to obtain the feature information after aggregating the information of each neighboring node and connected edge. This feature is multiplied by the 62*1 convolutional kernel of the attention layer. To achieve the nonlinear mapping capability of the network, the ReLU activation function is used to ensure that the output of the graph filtering layer is non-negative. Then, it passes through the Flatten layer, the first fully connected layer (10 62*1 convolutional kernels), and the second fully connected layer (3 10*1 convolutional kernels). Finally, the result is output, and the Softmax function is used to predict the expected class label information of the input EEG features.

[0043] This invention's model fully leverages the powerful information integration capabilities of graph convolution in the graph convolutional layer. Assuming uniform node features, the edge features connected to a channel node are convolved and concentrated onto that channel node. This aggregated node feature is then multiplied by the corresponding attention value and flows to the next layer. During training, the faster a node's features change across different categories, the larger the attention value corresponding to that node becomes after multiple training iterations. For example, if a node n0's value fluctuates between 1 and 100 in both class A and class B, with similar value distributions, then this node has almost no discriminative power for this classification task and its importance is very low. After multiple training iterations, its attention value will tend towards a random distribution, generally becoming a small absolute value, thus minimizing the loss when misclassified. Conversely, if another node n1 stably fluctuates around 0 in class A and around 1 in class B, then this node has good discriminative power for this task, and its attention value will become increasingly larger after multiple training iterations, leading to a larger loss when misclassified. After multiple iterations of training, these attention points are extracted and sorted in descending order according to their absolute values. This results in the importance ranking of each node. The Top K nodes are set as core nodes, and the edges connecting these core nodes constitute the core network.

[0044] The second objective of this invention is to provide a method for classifying and recognizing emotions from electroencephalogram (EEG) signals, including an extracted core network and a classifier.

[0045] Preferably, if the classifier is any one of SVM, KNN, or RF classifiers, then the input to the classifier is the core network extracted above. For the three machine learning classifiers SVM, KNN, and RF, the core network with shape k*k after extraction is used for training and classification. SVM uses a linear kernel function linear, and KNN sets n_neighbors to 5.

[0046] Preferably, if the classifier is a PGCNN deep learning classifier, the input of the classifier uses the DE of the core node of each sample as the point feature and the SCC of the input core network as the edge feature, which together constitute the graph feature signal input. For the PGCNN deep learning classifier, the input consists of the extracted DE node signal of shape k*b (b is the number of frequency bands) and the SCC edge signal (network) of shape k*k, which together constitute the graph signal input. The model of this PGCNN deep learning classifier consists of a series of gcn1 graph convolutional layer, pool1 graph pooling layer, dense1 fully connected layer and dense2 fully connected layer. The gcn1 graph convolutional layer has parameters of k*k and 10 b*1 convolutional kernels, with a graph convolution order of 2. The pool1 graph pooling layer uses SAGPooling with a ratio of 0.5. The dense1 and dense2 fully connected layers have parameters of 10 150*1 convolutional kernels and 3 10*1 convolutional kernels, respectively.

[0047] Preferably, if the classifier is a DGCNN deep learning classifier, the input of the classifier uses the core node DE of the sample as the point feature, and the average value of the functionally related edge feature SCC as the learnable dynamic matrix matrixA as the initial edge feature, because it approximately represents a certain intrinsic relationship between the interconnections between brain channels. The two constitute the graph feature signal input. For the DGCNN deep learning classifier, the extracted DE node signal with shape k*b (b is the number of frequency bands) is input first, and then the mean of the SCC core network in the training set is calculated as the initial value of matrixA in the model. The two constitute the graph signal. This DGCNN consists of a series of gcn1 graph convolutional layers, flattening layers, dense1 fully connected layers, and dense2 fully connected layers. The parameters of its gcn1 graph convolutional layer are k*k and 10 b*1 convolutional kernels, with a graph convolution order of 3. The parameters of the dense1 and dense2 fully connected layers are 10 300*1 convolutional kernels and 3 10*1 convolutional kernels, respectively.

[0048] A third object of the present invention is to provide a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method described thereon.

[0049] A fourth object of the present invention is to provide a computing device, including a memory and a processor, wherein the memory stores executable code, and the processor executes the executable code to implement the method described herein.

[0050] The beneficial effects of this invention are:

[0051] 1. This invention can be used to find the core nodes and core networks of brain networks in electroencephalogram (EEG) signals, providing assistance for specific analysis of EEG locations.

[0052] 2. The method of this invention reduces the number of nodes required for subsequent EEG signal processing, and reduces the computational complexity of classification and analysis.

[0053] 3. The method of the present invention improves the accuracy of EEG recognition.

[0054] 4. This invention uses the A-GCN model, which eliminates the need for manual extraction of the core information. Instead, it utilizes the complex information mining capabilities of deep learning to automatically extract the core information, thus improving flexibility. Attached Figure Description

[0055] Figure 1 This is a training framework diagram of the A-GCN attention map convolutional neural network model of the present invention;

[0056] Figure 2 This is a schematic diagram of the deep learning classifier PGCNN used in this invention;

[0057] Figure 3 This is a schematic diagram of the deep learning classifier DGCNN used in this invention;

[0058] Figure 4 The bar chart shows the accuracy results of the three machine learning classifiers used in this invention for different channel selection methods in five single-band and k=30 scenarios.

[0059] Figure 5 The bar chart shows the accuracy results of the five classifiers used in this invention under different channel selection methods in the gamma band and when k=10, 20, 30.

[0060] Figure 6 The bar chart shows the accuracy results of the two deep learning classifiers used in this invention for the channels selected by A-GCN in the full bandwidth k=10, 20, 30, 40, 50, 62.

[0061] Specific implementation method

[0062] The method for extracting the core network of EEG signals based on A-GCN according to the present invention will be described in detail below with reference to the accompanying drawings.

[0063] Based on the characteristics of brain networks in electroencephalogram (EEG) signals, this invention proposes an automatic core network extraction method, comprising the following steps:

[0064] Step (1). Preprocessing of EEG signals

[0065] This invention utilizes a single-subject experiment on the SEED dataset to validate the effectiveness of the three-category emotion classification model. Fifteen participants were involved in the experiment, with 15 different film clips played in each experiment to evoke corresponding emotions. These film clips represented three different emotion types: happiness, neutrality, and sadness. The participants wore a 62-channel ESI NeuroScan System to record the EEG signals generated during the film viewing in real time, with a sampling frequency of 1000Hz. To reduce storage space and computational load, some artifact noise was removed, and the EEG signals were downsampled to 200Hz.

[0066] 1-1 Data Format Standardization: In Figure 1 The paper demonstrates the process of standardizing the data format of EEG signals. First, bandpass filtering is performed on the 62-channel EEG data. Using Matlab software, the signal of each channel of the original signal is divided into five frequency bands: delta (1-4Hz), theta (4-8Hz), alpha (8-14Hz), beta (14-31Hz), and gamma (31-50Hz). Then, the channel dimension is organized into a 1D matrix of shape 62 according to the electrode distribution of the actual EEG signal acquisition. Finally, the EEG data is transformed into a 2D tensor data of shape 62*5 in both the channel dimension and the frequency band dimension.

[0067] 1-2 Data Segmentation and Integration: The EEG signal data obtained in step 1-1 was sliced ​​and processed using a 1-second non-overlapping Hanning time window for sliding window operation, resulting in a series of EEG signal segments with a shape of 200*62*5. The DE (Depth Analysis) was calculated on these original EEG signal segment samples, transforming them into node feature samples with a shape of 62*5. Similarly, the SCC (Side Correlation Conformity) was calculated on these EEG signal segment samples according to channel correlation, resulting in edge feature samples with a shape of 62*62*5. After data segmentation, for the SEED dataset, 9 out of 15 trials from each of the 15 participants were used as the training set, and the remaining 6 were used as the test set.

[0068] Step 2: Build the A-GCN model network: including a graph signal generation layer, a GCN graph convolutional neural network layer, an attention layer, a flattening layer, and two fully connected layers (e.g., ...). Figure 1 ).

[0069] The SCC (Side Cornering Convolutional Vector) samples of edge signal features in a certain frequency band are input into the A-GCN model network. To construct the graph signal and ensure that all information comes only from the edge signals, the graph signal generation layer generates a node feature vector with a dimension and value of 1 for each node of each sample. This node signal feature and the input edge signal feature are used to construct the graph signal feature according to the corresponding channels. The graph signal undergoes special convolution calculations through the GCN graph convolutional neural network layer. It treats the input of each node as a signal, and before performing the convolution operation, it uses the eigenvectors of the transposed normalized Laplacian matrix to perform a Fourier transform on the node signal. After the convolution is completed, it is transformed back using the eigenvectors of the normalized Laplacian matrix. In effect, it uses the information aggregation of the node, its neighboring nodes, and the connected edges. After exiting the GCN graph convolutional layer, the resulting feature vector has a shape of 62*1, similar to the shape of the node features. In the Attention layer, this feature vector is multiplied by the corresponding 62*1 attention vector at its position, resulting in a feature vector with the same shape. This vector is then activated by the ReLU activation function and fed into the Flatten layer, followed by two Dense fully connected layers for classification output. The softmax function is then used to predict the label classification. After multiple rounds of iterative training, the parameters in the model tend to stabilize. The attention parameters of the Attention layer are extracted and sorted in descending order of their absolute values. The top k parameters are selected as the TopK core nodes. These core nodes and the edges connecting them (the core network) are retained; the core network has a shape of k*k. Therefore, after the A-GCN model training is completed, the core nodes and core network for each subject's five frequency bands are obtained.

[0070] Step 3: For the three traditional machine learning classifiers, SVM, KNN, and RF, since they are suitable for single-feature classification and only have a single frequency band mode, only the obtained core network, i.e., the SCC correlation matrix of shape k*k, is input. However, for the PGCNN deep learning classifier (such as...), ... Figure 2 In single-band mode, its nodes use the point features corresponding to the channels of the frequency band, while its connectivity edges use the functionally relevant features of the frequency band. That is, the graph signal is composed of the DE node features corresponding to the input core node samples and the SCC edge features of the core network samples. For full-band mode, its edge features can only be 1-dimensional in the edge connectivity dimension. Therefore, the core node and core network of the best-performing gamma frequency band are selected, and the full-band channel dimension of the core node is taken as 5-dimensional DE (i.e., shape k*5), and the edge features are SCC with a 1-dimensional edge connectivity dimension (i.e., shape k*k). For DGCNN deep learning classifiers (such as...), Figure 3Similar to PGCNN, in single-band mode, the DE entropy node features corresponding to the core nodes of the input frequency band are used as the node features of the graph signal. The edge features of the graph signal are the average of the k*k SCC edge features of the core network of each sample in that frequency band, calculated from the training set. The shape is also k*k, and this average is used as the initial value of the common connected edge matrixA. MatrixA is a dynamically changing, learnable generated edge matrix used to learn some potential association information between brain channel nodes. For full-band mode, the core nodes and core network of the gamma frequency band are taken, and the full-band 5-dimensional DE on the core nodes is taken. The edge features are also taken as the average of the k*k SCC samples of the gamma frequency band core network as the initial value of matrixA. Then, the above five classifiers are trained.

[0071] Step 4: The performance evaluation results of the present invention on the SEED dataset are obtained by using the trained SVM, KNN, RF, PGCNN and DGCNN models to predict the labels of the samples in the test set, and the classification accuracy of the emotional EEG signals of 15 subjects is calculated. Finally, the average value is calculated.

[0072] This invention compares the original 62 full-channel signal with network and graph signals composed of k randomly selected channels and k key channels selected by A-GCN under multiple classifiers, and performs comparisons in different frequency bands. Experimental results show that when k takes different values ​​of 10, 20, and 30 (see... Figure 5 The core network constructed using k key channels (core nodes) selected by the A-GCN model significantly outperformed the core network constructed using k randomly selected channels across all frequency bands and classifiers. The most significant improvement was observed when k=30, with the full-band DGCNN classifier achieving an accuracy of 93.05%, even surpassing the 91.64% accuracy of the original 62-channel classifier by 1.41% (see [link to DGCNN model]). Figure 6 As the value of k decreases, the classification results of the core nodes and core networks selected by the A-GCN model decrease much more slowly under various classifiers than those obtained through random channel selection. This is especially true when k=10, where it still maintains high accuracy, demonstrating significantly higher classification performance than random channel selection. Furthermore, in the comparison of different single frequency bands (see...),... Figure 4The gamma band achieved the best results across all classifiers, followed by the bata band, while the theta and delta bands yielded poorer results. This indicates a stronger connection between emotion and higher-frequency EEG signals. This explains why our PGCNN and DGCNN deep learning classifiers used the gamma band to select core nodes and core networks when taking full-frequency features for node features. Therefore, the core network identified in this invention is not only suitable for classification prediction of single edge signal features (network), but also plays an important role in classification prediction of multiple feature collaborations (graph signals). It not only achieves core network extraction from the perspective of functional relevance, but also provides a feasible approach for core extraction of potential relationships between brain electrode channels.

[0073] Figure 5 The bar chart shows the accuracy results of the five classifiers used in this invention under different channel selection methods in the gamma band and when k=10, 20, 30.

[0074] Figure 6 The bar chart shows the accuracy results of the two deep learning classifiers used in this invention for the channels selected by A-GCN in the full bandwidth k=10, 20, 30, 40, 50, 62.

[0075] The A-GCN model designed in this invention automatically extracts representative signal features from EEG signals, thereby improving performance under various types of classifiers while reducing the number of channel electrode nodes. This not only greatly reduces the number of electrodes used and computational costs, but also improves recognition accuracy to a certain extent.

Claims

1. A method for extracting the core network of electroencephalogram (EEG) signals, characterized in that... Specifically, the steps include the following: Step (1): Preprocess EEG signals, label them with task categories, and construct a dataset; Step (2): Construct the Attention-Graph Convolutional Neural Network (A-GCN); The attention-graph convolutional neural network A-GCN takes pre-processed EEG signals of a certain frequency band as input and predicts task category labels as output. The attention-graph convolutional neural network A-GCN includes a cascaded graph signal generation layer, a GCN graph convolutional layer, an attention layer, a flatten layer, a first fully connected layer, and a second fully connected layer. The graph signal generation layer generates signals for each channel (n0, n1, ..., n). N-1 Generate a unified node signal x(x0, x1, ..., x...). N-1 The node signal x and the input edge signal W∈R N*N Together they form the feature vector of the graph signal; The GCN graph convolutional layer receives the graph signal feature vector output by the graph signal generation layer, and performs Fourier transform and Laplacian matrix eigenvalue decomposition on it before convolution calculation to obtain the feature information y after aggregating the information of each neighboring node and the connected edge. Specifically: The degree matrix D∈R is calculated from the edge signal W. N*N And obtain the Laplace matrix L = DW ∈ R N*N The characteristic decomposition form of L is UΛU T , where U=[u0,u1,....,u N-1 ]∈R N*N The columns form a Fourier basis, Λ=diag([λ0,λ1,...,λ N-1 ]) is a diagonal matrix; g θ The Fourier domain convolution function is y = g θ (L)x=g θ (UΛU T )x=U g θ (Λ)U T After convolution operation on x, we get y(y0, y1, ... y). N-1 ), y∈R N*1 ; The Attention layer receives the output of the GCN graph convolutional layer and multiplies it with an N*1 convolutional kernel to obtain the features processed by the attention mechanism. Specifically: Feature information y(y0, y1, ... y N-1 The convolution kernel A(a0, a1, ..., a) corresponding to the attention node N-1 Multiplication: Where A∈R N*1 , This represents element-wise multiplication; The Flatten layer receives the features output by the attention layer, which are activated by the ReLU function y = ReLU(y) and flatten the data. Then, the processed data is fully mapped through the first fully connected layer and the second fully connected layer to output the result, and the Softmax function is used to predict the label classification. Step (3): Train the Attention-Graph Convolutional Neural Network A-GCN in multiple rounds of iterations, and obtain the core node set C based on the attention of the Attention layer in the Attention-Graph Convolutional Neural Network A-GCN after training; Specifically The loss is calculated after comparing y with the actual data label class: Subsequently, tensor backpropagation is used for gradient descent to adjust the convolution kernel parameters. After multiple rounds of training iterations, the convolution kernel parameters A(a0, a1, ... a2) of the Attention layer are adjusted. N-1 The situation is trending towards stability; After multiple rounds of iterative training of the Attention-Graph Convolutional Neural Network (A-GCN), the final attentions in the Attention layer are sorted in descending order according to the absolute value of the elements in A. The top K final attentions are selected, and their corresponding original channels are used as core nodes C(c0, c1, c2, ..., c...). K-1 ); Step (4): Based on the core node C, retain the corresponding interconnected edges w. ij Where i, j∈(c0, c1, ... c K-1 Finally, the core network S∈R is obtained. K*K .

2. The method as described in claim 1, characterized in that... Step (1) specifically involves: 1-1 Data Format Standardization Based on the electrode positions of the collected EEG signals, the feature vectors of the total channels are recorded in 1D matrix format in order from left to right and from top to bottom, where N is the total number of channels. At the same time, the EEG signal of each channel is divided into 5 frequency bands to obtain EEG data of 2D tensor data with shape N*5. The five frequency bands mentioned are delta, theta, alpha, beta, and gamma; 1-2 Data Splitting and Cleaning The EEG signal obtained in step 1-1 is sliced, and a non-overlapping sliding window operation is performed with T as the time window to obtain a series of EEG signal segments with shape L*N*5, where L is the data length, L=T*w, and w is the sampling frequency; The differential entropy (DE) and Spearman rank correlation coefficient (SCC) are calculated for the sliced ​​EEG signal segments. The differential entropy (DE) is used as the node signal and the Spearman rank correlation coefficient (SCC) is used as the edge signal, resulting in a series of N*5 node signals and N*N*5 edge signals. At the same time, the sliced ​​EEG signal segments are labeled with their respective task categories, and finally a classification dataset is constructed.

3. The method as described in claim 1, characterized in that... The graph signal generation layer generates a node signal for each node, with a vector dimension and value of 1. This node signal, together with the input Spearman rank correlation coefficient (SCC) edge signal, constitutes the graph signal.

4. The method as described in claim 1, characterized in that... The convolution kernel size of the GCN graph convolution layer is 62*62 and 3*1, and the graph convolution order is 3.

5. The method as described in claim 1, characterized in that... The attention convolution kernels of the attention layer correspond to each node at a fixed position. There are 62 1*1 convolution kernels with an initial value of 1.

6. The method as described in claim 1, characterized in that... Flatten layers reduce multidimensional features activated by the ReLU function to one dimension.

7. The method as described in claim 1, characterized in that... The first fully connected layer and the second fully connected layer have 10 62*1 convolutional kernels and 3 10*1 convolutional kernels, respectively.

8. A brainwave signal emotion recognition and classification method based on the method of any one of claims 1-7, characterized in that... This includes the extracted core network and classifier.

9. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method of any one of claims 1-8.

10. A computing device comprising a memory and a processor, wherein the memory stores executable code, and the processor, when executing the executable code, implements the method of any one of claims 1-8.

Images