Electroencephalogram emotion recognition method and system based on adaptive multi-view graph neural network

Abstract

This invention relates to a method and system for EEG emotion recognition based on an adaptive multi-view graph neural network, belonging to the field of brain-computer interface and emotion computing technology. The method includes: dividing multi-channel EEG signals into continuous time windows, and using four adjacent time windows as temporal input samples; extracting multi-band differential entropy features of each time window as initial node features; fusing prior knowledge of electrode spatial proximity and brain biological symmetry to construct a basic matrix, and modulating and applying sparse constraints through a learnable attention mechanism to generate an individualized brain functional connectivity topology; designing a parallel bi-branch deep network, where a graph convolutional branch extracts global spatiotemporal features from the graph structure sequences corresponding to the four time windows, and a one-dimensional convolutional branch extracts and fuses local frequency-spatial features; and during training, comprehensively applying node-level domain adversarial and graph structure collaborative regularization to output the emotion category. This invention is beneficial for improving cross-subject recognition performance.

Description

Technical Field

[0001] This invention belongs to the field of brain-computer interface and emotion computing technology, specifically relating to a brainwave emotion recognition method and system based on an adaptive multi-view graph neural network. Background Technology

[0002] Emotion recognition has significant applications in human-computer interaction, mental health assessment, and neuroscience. Due to the advantages of electroencephalography (EEG), such as high temporal resolution, non-invasiveness, and difficulty in spoofing, EEG-based emotion recognition has become a research hotspot. In recent years, deep learning, especially graph neural networks (GNNs), has made significant progress in this field because it can effectively model the non-Euclidean topological relationships between multi-channel EEG signals. However, existing technologies still have the following prominent problems:

[0003] Insufficient individual adaptability in brain topology modeling: Most GNN methods use fixed adjacency matrices based on prior spatial distance, or fully data-driven but unconstrained dynamic adjacency matrices, which make it difficult to simultaneously take into account prior knowledge in neuroscience (such as brain region functional division) and differences in brain activity among individuals, resulting in brain functional maps that may not be accurate or stable.

[0004] Feature extraction is often limited to a single dimension: existing methods tend to focus on extracting only one type of feature. For example, standard GNNs primarily capture spatial topological features between channels, while CNNs excel at extracting local spatial and spectral patterns, and RNNs / LSTMs are adept at modeling temporal dynamics. However, emotional information is simultaneously contained in multiple dimensions of EEG signals, including spatial coordination, spectral energy, and temporal evolution, making it difficult for a single model to capture them all comprehensively.

[0005] The severe lack of cross-subject generalization ability (domain shift) is a core bottleneck hindering the practical application of EEG emotion recognition. Individual physiological and psychological differences lead to significant variations in the distribution of EEG signals (including features and brain functional connectivity patterns) among different subjects. Although some studies have attempted to introduce domain adaptation techniques, most methods only align at the overall sample feature level, ignoring the fact that different brain regions (corresponding to nodes in the figure) may have different domain shift characteristics. This results in coarse alignment granularity and limited effectiveness.

[0006] In summary, there is currently a lack of an integrated solution that can adaptively construct individualized brain function maps, integrate multi-dimensional features, and perform fine-grained cross-subject domain adaptation. Summary of the Invention

[0007] The purpose of this invention is to overcome the above-mentioned defects of the prior art and provide a brainwave emotion recognition method and system based on an adaptive multi-view graph neural network.

[0008] To achieve the above objectives, the technical solution of the present invention is: a brainwave emotion recognition method based on an adaptive multi-view graph neural network, comprising:

[0009] EEG feature extraction steps: Preprocess the multi-channel EEG signals and divide each EEG signal into multiple consecutive time windows; combine four adjacent time windows in chronological order to form a time-series input sample, i.e., a time window sequence; for each channel within each time window in the time window sequence, extract its differential entropy features in multiple preset frequency bands to construct the initial node feature vector corresponding to each channel. The initial node feature vectors of all channels together constitute the initial node feature matrix corresponding to that time window; one time-series input sample corresponds to four initial node feature matrices.

[0010] The adaptive brain functional topology map construction steps are as follows: taking each EEG acquisition channel as a graph node, a sparse adjacency matrix is ​​dynamically generated based on the initial node feature matrix to characterize the functional connectivity of the current EEG sample; in this step, a learnable attention modulation mechanism is used to weight and modulate a fixed basis matrix that integrates electrode spatial proximity and brain biological symmetry priors, and sparsity constraints are applied to the modulation result to obtain the sparse adjacency matrix; in particular, four sparse adjacency matrices are generated for each of the four time windows in a temporal input sample.

[0011] The multi-view feature parallel extraction and fusion steps are as follows: The initial node feature matrix and sparse adjacency matrix are input into a parallel two-branch deep neural network for feature extraction. The first feature extraction branch is a graph convolution branch, which defines a graph structure using the sparse adjacency matrix. It performs graph convolution feature extraction on the four time windows of the temporal input sample to obtain four graph-level feature vectors. These four graph-level feature vectors are then input into the temporal modeling module in chronological order to capture the dynamic evolution of emotional states and output global spatiotemporal fusion features. The second feature extraction branch is a one-dimensional convolutional neural network branch. After reconstructing the initial node feature matrix into a one-dimensional sequence, it extracts local frequency domain-electrode channel features between channels through one-dimensional deep convolution operations and outputs local frequency-spatial features. Specifically, local frequency-spatial features are extracted for the four time windows using parameter-sharing convolution branches, and the local frequency-spatial features of the four time windows are converged to obtain local frequency-spatial features. Finally, the global spatiotemporal fusion features and local frequency-spatial features are concatenated and fused to obtain a multi-view fusion feature representation for emotion classification.

[0012] Model optimization and regularization training steps: During model training, a comprehensive loss function is minimized; the comprehensive loss function is composed of a weighted sum of standard sentiment classification loss, node-level domain adversarial regularization loss, and graph structure domain collaborative regularization loss; wherein, the node-level domain adversarial regularization aims to perform adversarial training by configuring a domain discriminator for each node in the graph, so as to force the feature distribution of each graph node learned by the feature extraction network to be difficult to distinguish among different subjects; the graph structure domain collaborative regularization aims to constrain the sparse adjacency matrices corresponding to different subject samples to tend to be consistent in statistical properties or distribution;

[0013] Emotional state classification steps: The EEG signal to be identified is processed through the above steps, and its multi-view fusion feature representation is input into the classifier trained by the model optimization and regularization training steps, and the corresponding emotion category result is output.

[0014] Furthermore, in the adaptive brain functional topology map construction step, the learnable attention modulation mechanism is specifically implemented as follows:

[0015] The initial node feature matrix is ​​mapped through a neural network to generate the query matrix Q and the key matrix K, respectively.

[0016] Calculate the product of the query matrix Q and the transpose of the key matrix K, and then perform scaling and softmax normalization to obtain an attention weight matrix that reflects the strength of functional connections between channels. ;

[0017] attention weight matrix Element-wise multiplication with the fixed fundamental matrix B is performed, i.e., the Hadamard product operation is executed. = ⊙ This enables the modulation of prior spatial connections by functional connection attention; wherein, the fixed basis matrix B is a fixed basis matrix that integrates the priors of electrode spatial proximity and brain biological symmetry.

[0018] Furthermore, the application of sparsity constraints may be carried out in one or a combination of the following ways:

[0019] For the modulated matrix Each element in the array is subjected to a threshold-based truncation function, and elements that are less than a preset threshold τ are set to zero.

[0020] During training, the modulated matrix is ​​added to the loss function. Alternatively, the L1 norm regularization term corresponding to the sparse adjacency matrix can be used to promote the sparsity of the adjacency matrix.

[0021] Furthermore, the anti-oversmoothing design in the graph convolution branch refers to including an adjacency matrix enhancement operation and a residual connection operation in the computation process of each graph convolution block. The adjacency matrix enhancement operation adds the identity matrix with learnable weights to the sparse adjacency matrix to obtain an enhanced adjacency matrix, which is used for graph convolution computation to preserve the node's own features during information propagation. The residual connection operation adds the input features of the current graph convolution block to the features obtained after graph convolution operation via the enhanced adjacency matrix to obtain the output features of the current graph convolution block.

[0022] Furthermore, the temporal modeling module employs a bidirectional long short-term memory network or a Transformer encoder based on a self-attention mechanism to model the graph-level feature sequences extracted by the graph convolutional branches and arranged in temporal order, in order to capture the temporal evolution pattern of emotional states.

[0023] Furthermore, the specific implementation of the node-level domain adversarial regularization loss is as follows:

[0024] At the end of the graph convolution branch, each node in the graph is equipped with a lightweight domain discriminator subnetwork. The lightweight domain discriminator subnetwork is a neural network containing 1-2 fully connected layers. The output dimension of the fully connected layers is 2, corresponding to the binary classification of the source domain and the target domain.

[0025] During training, the gradient signal from the domain discriminator subnetwork is backpropagated to the feature extractor through the gradient inversion layer;

[0026] The optimization goal of the feature extractor is to enable the features of all nodes to deceive their respective domain discriminator subnetworks, while the optimization goal of each domain discriminator subnetwork is to accurately determine the source domain of its corresponding node features. This adversarial game achieves node-level feature distribution alignment.

[0027] Furthermore, the graph structure domain collaborative regularization loss includes two components: intra-graph consistency loss and inter-graph distribution alignment loss.

[0028] Intragraph consistency loss: Calculate the difference between the sparse adjacency matrices corresponding to different EEG samples within the same subject, and minimize this difference to improve the stability of the subject's brain functional topology; Intragraph consistency loss is obtained by calculating the mean of the Frobenius norm between different sparse adjacency matrices within the same subject.

[0029] Inter-graph distribution alignment loss: Calculate the maximum mean difference between the sparse adjacency matrix sets of source domain subjects and target domain subjects in the Hilbert space of the regeneration kernel, and minimize this difference to bring the distribution of brain functional connectivity patterns between different subjects closer together.

[0030] The present invention also provides an EEG emotion recognition system based on an adaptive multi-view graph neural network, comprising: an EEG feature extraction module, an adaptive brain function topology graph construction module, a multi-view feature parallel extraction and fusion module, a model optimization and regularization training module, and an emotion state classification module; each module is used to perform the steps described above.

[0031] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the EEG emotion recognition method based on an adaptive multi-view graph neural network as described above.

[0032] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the EEG emotion recognition method based on an adaptive multi-view graph neural network as described above.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1. Through a mechanism combining attention modulation and sparse constraints, we adaptively construct brain functional topologies that better reflect the characteristics of individual neural activity.

[0035] 2. Design a dual-branch deep network architecture that utilizes the advantages of graph convolution and conventional convolution to extract global spatial topological features and local frequency-space features of EEG signals in parallel, and obtains a more comprehensive emotion representation through deep fusion.

[0036] 3. A dual regularization strategy of node-level domain adversarial training and graph structure consistency constraint is introduced to align the EEG feature distribution and brain network patterns of different subjects at a fine-grained level, which significantly improves the recognition accuracy and robustness of the model in subject-independent scenarios. Attached Figure Description

[0037] Figure 1 This is an overall flowchart of an EEG emotion recognition method provided in an embodiment of the present invention;

[0038] Figure 2 This is a detailed structural schematic diagram of the adaptive brain function map construction module in one embodiment of the present invention;

[0039] Figure 3 This is a schematic diagram of the structure of a multi-branch deep feature extraction network in one embodiment of the present invention;

[0040] Figure 4 This is a schematic diagram of a node-level domain adversarial training strategy (NodeDAT) in one embodiment of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are for illustrative purposes only and are not intended to limit the invention.

[0042] This invention provides a brainwave emotion recognition method based on an adaptive multi-view graph neural network, comprising:

[0043] EEG feature extraction steps: Preprocess the multi-channel EEG signals and divide each EEG signal into multiple consecutive time windows; combine four adjacent time windows in chronological order to form a time-series input sample, i.e., a time window sequence; for each channel within each time window in the time window sequence, extract its differential entropy features in multiple preset frequency bands to construct the initial node feature vector corresponding to each channel. The initial node feature vectors of all channels together constitute the initial node feature matrix corresponding to that time window; one time-series input sample corresponds to four initial node feature matrices.

[0044] The adaptive brain functional topology map construction steps are as follows: taking each EEG acquisition channel as a graph node, a sparse adjacency matrix is ​​dynamically generated based on the initial node feature matrix to characterize the functional connectivity of the current EEG sample; in this step, a learnable attention modulation mechanism is used to weight and modulate a fixed basis matrix that integrates electrode spatial proximity and brain biological symmetry priors, and sparsity constraints are applied to the modulation result to obtain the sparse adjacency matrix; in particular, four sparse adjacency matrices are generated for each of the four time windows in a temporal input sample.

[0045] The multi-view feature parallel extraction and fusion steps are as follows: The initial node feature matrix and sparse adjacency matrix are input into a parallel two-branch deep neural network for feature extraction. The first feature extraction branch is a graph convolution branch, which defines a graph structure using the sparse adjacency matrix. It performs graph convolution feature extraction on the four time windows of the temporal input sample to obtain four graph-level feature vectors. These four graph-level feature vectors are then input into the temporal modeling module in chronological order to capture the dynamic evolution of emotional states and output global spatiotemporal fusion features. The second feature extraction branch is a one-dimensional convolutional neural network branch. After reconstructing the initial node feature matrix into a one-dimensional sequence, it extracts local frequency domain-electrode channel features between channels through one-dimensional deep convolution operations and outputs local frequency-spatial features. Specifically, local frequency-spatial features are extracted for the four time windows using parameter-sharing convolution branches, and the local frequency-spatial features of the four time windows are converged to obtain local frequency-spatial features. Finally, the global spatiotemporal fusion features and local frequency-spatial features are concatenated and fused to obtain a multi-view fusion feature representation for emotion classification.

[0046] Model optimization and regularization training steps: During model training, a comprehensive loss function is minimized; the comprehensive loss function is composed of a weighted sum of standard sentiment classification loss, node-level domain adversarial regularization loss, and graph structure domain collaborative regularization loss; wherein, the node-level domain adversarial regularization aims to perform adversarial training by configuring a domain discriminator for each node in the graph, so as to force the feature distribution of each graph node learned by the feature extraction network to be difficult to distinguish among different subjects; the graph structure domain collaborative regularization aims to constrain the sparse adjacency matrices corresponding to different subject samples to tend to be consistent in statistical properties or distribution;

[0047] Emotional state classification steps: The EEG signal to be identified is processed through the above steps, and its multi-view fusion feature representation is input into the classifier trained by the model optimization and regularization training steps, and the corresponding emotion category result is output.

[0048] The following are specific implementation examples of the present invention.

[0049] like Figure 1-4 As shown, a brainwave emotion recognition method based on an adaptive multi-view graph neural network includes the following steps:

[0050] S101. Data Preparation and Preprocessing: The publicly available dataset SEED was used, containing 62 channels of EEG signals at a sampling rate of 200Hz. Bandpass filtering of 0.5-50Hz was applied to the raw signals to remove high-frequency noise and power line interference. Independent component analysis (ICA) was used to remove electrooculography (EOG) artifacts. Each experimental data segment was segmented using a 1-second window with a 0.5-second overlap. The four adjacent time windows obtained from the sliding segmentation were arranged chronologically to form a time window sequence as a temporal input sample. Each time window in the time window sequence corresponds to an initial node feature matrix and a sparse adjacency matrix, and the four time windows share the same sentiment label. For each time window t (t=1,2,3,4) in the temporal input sample, an initial node feature matrix was calculated. (of shape [62,5]) and construct a sparse adjacency matrix. (Shape is [62, 62]), thus forming a feature sequence of length 4 { } and graph structure sequence { }

[0051] S102. Feature Extraction: For the signal within each time window, calculate its differential entropy (DE) features in five frequency bands: δ band [1,4) Hz, θ band [4,8) Hz, α band [8,14) Hz, β band [14,31) Hz, and γ band [31,50] Hz. For each time window, obtain a feature matrix of shape [62 electrode channels, 5 frequency bands], which serves as the initial features for the nodes. ∈ ,in =62, =5.

[0052] S103. Adaptive Brain Function Map Construction: Constructing the Basic Matrix B: Calculate the 3D Euclidean distance matrix Dist for the 62 electrodes in the standard brain model. For each node i, select its k nearest neighbors (k=8) and electrodes at symmetrical positions in the left and right hemispheres, setting the corresponding positions in B to 1, and the rest to 0. This yields a sparse 0-1 matrix B that reflects spatial proximity and biological symmetry.

[0053] Attention weight generation: The initial node features X are passed through a shared fully connected layer to generate a query matrix Q and a key matrix K. Attention scores are then calculated. , where d is the vector dimension. Perform softmax normalization row by row to obtain the attention weight matrix. .

[0054] Modulation and Sparsification: Computing Adaptive Adjacency Matrix = ⊙ , where ⊙ represents the Hadamard product (element-by-element multiplication). For Each element is activated by the ReLU function, and a threshold τ = 0.1 is set, setting values ​​less than τ to zero, thus obtaining a sparse, non-negative adjacency matrix.

[0055] S104. Configuration and Training of Multi-Branch Deep Networks:

[0056] Branch 1 (GCN-LSTM branch): The graph convolutional layer is configured with 2 blocks. Each block contains: an initial residual connection (adding the input features directly to the output of this layer), two graph convolutional layers (using Chebyshev polynomial approximation, order K=3), an identity mapping, and layer normalization and ReLU activation. The output dimensions are 128 and 256, respectively. For each time window t in the temporal input samples, the corresponding... and Inputting the GCN branch yields node features (shape [62, 256]), and global average pooling is performed on the node dimension to obtain graph-level features. (Shape is [1,256]). The graph-level features of four consecutive time windows are stacked in chronological order to form a feature sequence. (Shape [4,256]), input a bidirectional LSTM (128 hidden units) to extract temporal features, and use the output of the bidirectional LSTM as the global spatiotemporal features. (The shape is [1,256]).

[0057] Branch 2 (1D-CNN branch): For the four consecutive time windows corresponding to Branch 1, extract the initial feature matrix for each time window. , where 62 represents the number of electrode channels and 5 represents the number of frequency bands. The initial node feature matrix is ​​transposed in a preset manner, and the initial feature matrix of each time window is organized into an input form suitable for one-dimensional convolution along the electrode channel dimension, with 5 frequency bands as input channels and 62 electrode channels as the one-dimensional sequence length. The one-dimensional convolution is performed along the electrode channel dimension to extract the local joint response features of different frequency bands on each electrode channel. The four time windows are used to extract features using a parameter-shared 1D-CNN network, with the network structure as follows: Conv1D(64, kernel=3) -> ReLU -> MaxPool1D(2) -> Conv1D(128, kernel=3) -> ReLU -> GlobalAvgPool1D(). After the above network processing, each time window outputs a local frequency domain-electrode channel feature. ([1,128]), and then average pooling is performed to fuse the features corresponding to the four time windows: The output features of branch two are obtained. .

[0058] Feature fusion and classification: and spliced ​​together The features of [1, 384] will The input is further fused into a fully connected layer (output dimension 128, ReLU activation), and finally outputs the sentiment probability through a 3-dimensional Softmax classification layer.

[0059] Regularization strategies: ①NodeDAT: Implemented on the node features ([62, 256]) output by the last GCN Block. A small domain discriminator (two-layer fully connected network) is constructed for each of the 62 nodes. During training, a subject is randomly selected as the "target domain" in each iteration, and the rest are the "source domains". The sum of the domain classification losses on the 62 nodes is calculated, and adversarial training is performed through a gradient reversal layer. λ1 is set to 0.1. ②GraphDCL: Calculates the domain classification loss of all samples in the current batch. Matrix, calculating the inter-sample distance in the source domain The mean of the Frobenius norm difference is used as the consistency loss, and the differences between the source and target domains are calculated. The distributional differences (such as MMD distance) are used as the alignment loss. λ2 is set to 0.05.

[0060] Training details: The Adam optimizer was used with an initial learning rate of 0.0005, a batch size of 16, and 100 training epochs. Early stopping was performed using the accuracy on the validation set.

[0061] S105. Evaluation: Leave-one-out-of-participant cross-validation (LOSO) was used on the SEED dataset, which is a standard protocol for evaluating cross-participant generalization ability.

[0062] Experimental results: The method of this invention significantly outperforms the comparative methods in the cross-subject-independent recognition accuracy (e.g., an improvement of 3%-5% compared to the best baseline), while also maintaining a leading or comparable level in the in-subject-dependent task, verifying that it improves generalization ability without compromising basic performance.

[0063] The present invention also provides an EEG emotion recognition system based on an adaptive multi-view graph neural network, comprising: an EEG feature extraction module, an adaptive brain function topology graph construction module, a multi-view feature parallel extraction and fusion module, a model optimization and regularization training module, and an emotion state classification module; each module is used to perform the steps described above.

[0064] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the EEG emotion recognition method based on an adaptive multi-view graph neural network as described above.

[0065] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the EEG emotion recognition method based on an adaptive multi-view graph neural network as described above.

[0066] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. An electroencephalogram emotion recognition method based on an adaptive multi-view graph neural network, characterized in that, include: EEG feature extraction steps: Preprocess the multi-channel EEG signals and divide each EEG signal into multiple consecutive time windows; combine four adjacent time windows in chronological order to form a time-series input sample, i.e., a time window sequence; for each channel within each time window in the time window sequence, extract its differential entropy features in multiple preset frequency bands to construct the initial node feature vector corresponding to each channel. The initial node feature vectors of all channels together constitute the initial node feature matrix corresponding to that time window; one time-series input sample corresponds to four initial node feature matrices. The adaptive brain functional topology map construction steps are as follows: taking each EEG acquisition channel as a graph node, a sparse adjacency matrix is ​​dynamically generated based on the initial node feature matrix to characterize the functional connectivity of the current EEG sample; in this step, a learnable attention modulation mechanism is used to weight and modulate a fixed basis matrix that integrates electrode spatial proximity and brain biological symmetry priors, and sparsity constraints are applied to the modulation result to obtain the sparse adjacency matrix; in particular, four sparse adjacency matrices are generated for each of the four time windows in a temporal input sample. The multi-view feature parallel extraction and fusion steps are as follows: The initial node feature matrix and sparse adjacency matrix are input into a parallel two-branch deep neural network for feature extraction. The first feature extraction branch is a graph convolution branch, which defines a graph structure using the sparse adjacency matrix. It performs graph convolution feature extraction on the four time windows of the temporal input sample to obtain four graph-level feature vectors. These four graph-level feature vectors are then input into the temporal modeling module in chronological order to capture the dynamic evolution of emotional states and output global spatiotemporal fusion features. The second feature extraction branch is a one-dimensional convolutional neural network branch. After reconstructing the initial node feature matrix into a one-dimensional sequence, it extracts local frequency domain-electrode channel features between channels through one-dimensional deep convolution operations and outputs local frequency-spatial features. Specifically, local frequency-spatial features are extracted for the four time windows using parameter-sharing convolution branches, and the local frequency-spatial features of the four time windows are converged to obtain local frequency-spatial features. Finally, the global spatiotemporal fusion features and local frequency-spatial features are concatenated and fused to obtain a multi-view fusion feature representation for emotion classification. Model optimization and regularization training steps: During model training, a comprehensive loss function is minimized; the comprehensive loss function is composed of a weighted sum of standard sentiment classification loss, node-level domain adversarial regularization loss, and graph structure domain collaborative regularization loss; wherein, the node-level domain adversarial regularization aims to perform adversarial training by configuring a domain discriminator for each node in the graph, so as to force the feature distribution of each graph node learned by the feature extraction network to be difficult to distinguish among different subjects; the graph structure domain collaborative regularization aims to constrain the sparse adjacency matrices corresponding to different subject samples to tend to be consistent in statistical properties or distribution; Emotional state classification steps: The EEG signal to be identified is processed through the above steps, and its multi-view fusion feature representation is input into the classifier trained by the model optimization and regularization training steps, and the corresponding emotion category result is output.

2. The electroencephalogram emotion recognition method based on the adaptive multi-view graph neural network according to claim 1, wherein, In the adaptive brain functional topology map construction step, the learnable attention modulation mechanism is specifically implemented as follows: The initial node feature matrix is ​​mapped through a neural network to generate the query matrix Q and the key matrix K, respectively. Calculate the product of the query matrix Q and the transpose of the key matrix K, and then perform scaling and softmax normalization to obtain an attention weight matrix that reflects the strength of functional connections between channels. ; attention weight matrix Element-wise multiplication with the fixed fundamental matrix B is performed, i.e., the Hadamard product operation is executed. = ⊙ This enables the modulation of prior spatial connectivity through functional connectivity attention; wherein, the fixed basis matrix B is a fixed basis matrix that integrates prior knowledge of electrode spatial proximity and brain biological symmetry. It is an adaptive adjacency matrix.

3. The EEG emotion recognition method based on adaptive multi-view graph neural network according to claim 2, characterized in that, The applying the sparsity constraint comprises: modulating the matrix A threshold-based truncation function is performed to set elements less than a preset threshold τ to zero to obtain a sparse adjacency matrix.

4. The EEG emotion recognition method based on adaptive multi-view graph neural network according to claim 2 or 3, characterized in that, In the training process, L1 norm regular term corresponding to the modulated matrix or sparse adjacency matrix is added in the loss function to promote the sparsification of the adjacency matrix.

5. The electroencephalogram emotion recognition method based on the adaptive multi-view graph neural network according to claim 1, characterized in that, The anti-oversmoothing design in the graph convolution branch includes: in each graph convolution block, multiplying the identity matrix by the learnable weight coefficients and adding it to the sparse adjacency matrix to obtain the enhanced adjacency matrix; performing graph convolution operation on the input features of the current graph convolution block based on the enhanced adjacency matrix; and adding the input features of the current graph convolution block to the features obtained from the graph convolution operation to obtain the output features of the current graph convolution block.

6. The electroencephalogram emotion recognition method based on the adaptive multi-view graph neural network according to claim 1, characterized in that, The temporal modeling module employs a bidirectional long short-term memory network or a Transformer encoder based on a self-attention mechanism to model the graph-level feature sequences extracted by the graph convolutional branches and arranged in temporal order, in order to capture the temporal evolution pattern of emotional states.

7. The electroencephalogram emotion recognition method based on the adaptive multi-view graph neural network according to claim 1, characterized in that, The specific implementation method of the node-level domain adversarial regularization loss is as follows: At the end of the graph convolution branch, each node in the graph is equipped with a lightweight domain discriminator subnetwork. The lightweight domain discriminator subnetwork is a neural network containing 1-2 fully connected layers. The output dimension of the fully connected layers is 2, corresponding to the binary classification of the source domain and the target domain. During training, the gradient signal from the domain discriminator subnetwork is backpropagated to the feature extractor through the gradient inversion layer; The optimization goal of the feature extractor is to enable the features of all nodes to deceive their respective domain discriminator subnetworks, while the optimization goal of each domain discriminator subnetwork is to accurately determine the source domain of its corresponding node features. This adversarial game achieves node-level feature distribution alignment.

8. The EEG emotion recognition method based on adaptive multi-view graph neural network according to claim 1, characterized in that, The graph structure domain collaborative regularization loss includes two components: intra-graph consistency loss and inter-graph distribution alignment loss. Intragraph consistency loss: Calculate the difference between the sparse adjacency matrices corresponding to different EEG samples within the same subject and minimize this difference to improve the stability of the subject's brain functional topology; Intragraph consistency loss is obtained by calculating the mean of the Frobenius norm between different sparse adjacency matrices within the same subject. Inter-graph distribution alignment loss: Calculate the maximum mean difference between the sparse adjacency matrix sets of source domain subjects and target domain subjects in the Hilbert space of the regeneration kernel, and minimize this difference to bring the distribution of brain functional connectivity patterns between different subjects closer together.

9. An electroencephalogram emotion recognition system based on an adaptive multi-view graph neural network, characterized in that, include: The system comprises an EEG feature extraction module, an adaptive brain function topology map construction module, a multi-view feature parallel extraction and fusion module, a model optimization and regularization training module, and an emotion state classification module; each module is used to execute the method steps as described in any one of claims 1-8.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by the processor, it implements the EEG emotion recognition method based on an adaptive multi-view graph neural network as described in any one of claims 1-8.

Images