An electroencephalogram signal emotion recognition method and system based on hierarchical dynamic mask

Abstract

The application discloses a layered dynamic mask-based electroencephalogram signal emotion recognition method and system, the method comprises the following steps: establishing a graph structure for the input electroencephalogram signal to generate a graph representation of the electroencephalogram signal in multiple specified frequency bands; extracting and fusing features of the graph representation of the electroencephalogram signal in each specified frequency band by using a layered dynamic mask graph convolutional network; the layered dynamic mask graph convolutional network comprises a multistage graph convolutional network (GCN), and a channel relationship adjacency matrix is dynamically updated by using a gradient descent method; and finally, the fused features extracted by the layered dynamic mask graph convolutional network are classified by using a classification module to obtain a final emotion classification result. The application aims to guide the alignment of the conditional distribution of the source domain and target domain data in the domain adaptation process of the multi-source domain adaptation model, realize different feature analysis and alignment across subjects and across sessions, and improve the accuracy and generalization performance of the electroencephalogram signal recognition across domains.

Description

Technical Field

[0001] This invention relates to the field of electroencephalogram (EEG) signal analysis technology, specifically to an EEG signal emotion recognition method and system based on hierarchical dynamic masking. Background Technology

[0002] The current EEG emotion recognition technology faces three major technical bottlenecks: (1) False connection interference: When traditional graph convolutional networks construct EEG functional connectivity maps, a large number of false connections caused by physical proximity rather than real neural activity are generated due to the volume conduction effect. For example, in the DGCNN method, the proportion of false connections between adjacent electrodes is as high as 62%, which seriously affects the model's capture of real emotion-related brain region interactions. (2) Lack of multi-scale features: Existing methods such as the V-IAG model only use single-layer graph convolution, which cannot capture local electrode activity and whole-brain network collaborative patterns at the same time. Existing literature experiments show that ignoring multi-scale features will reduce the cross-subject recognition accuracy by 12.7%. (3) Insufficient interpretability: Traditional methods such as GCB-Net lack quantitative analysis of decision-making basis, making it difficult to verify whether the model conforms to the laws of neuroscience. SEED-IV dataset tests show that the unverified interpretation scheme has insufficient matching degree with known emotional brain regions. Existing solutions such as Dynamic Graph Convolutional Network (DGCNN) and sparse constraint (Grop) have partially alleviated the above problems, but there are still problems such as incomplete suppression of false connections (residual redundant edges) and coarse feature fusion. Summary of the Invention

[0003] The technical problem to be solved by this invention is to provide a method and system for emotion recognition of EEG signals based on hierarchical dynamic masking, which addresses the above-mentioned problems in the prior art. This invention aims to guide the alignment of the data conditional distribution between the source domain and the target domain during the domain adaptation process of the multi-source domain adaptation model, realize the analysis and alignment of different features across subjects and sessions, and improve the accuracy and generalization performance of cross-domain EEG signal recognition.

[0004] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for emotion recognition based on hierarchical dynamic masking of EEG signals includes the following steps: S101, A graph structure is constructed for the input EEG signal to generate a graph representation of EEG signals in multiple specified frequency bands. In the graph representation of the EEG signal, nodes represent the EEG signal characteristics of the electrode channels, and edges represent the functional connections between the electrode channels; the channel relationship adjacency matrix is ​​initialized. ; S102, the graphical representations of EEG signals in each specified frequency band are used to extract fusion features using a hierarchical dynamic masked graph convolutional network. This hierarchical dynamic masked graph convolutional network includes a multi-level graph convolutional network (GCN) combined with a channel relation adjacency matrix. The importance score of each node in the graph representation of the EEG signal is calculated. Non-critical nodes are masked layer by layer using the importance score and different node masks as input to each level of the Graph Convolutional Network (GCN). Features extracted by the multi-level GCN are fused across layers to obtain the final extracted fused features. The channel adjacency matrix is ​​dynamically updated using gradient descent. ; S103 uses the classification module to classify the fused features extracted by the hierarchical dynamic mask image convolutional network to obtain the final emotion classification result.

[0005] Optionally, in step S102, the channel relationship adjacency matrix is ​​combined. The functional expression for calculating the importance score of each node in the graphical representation of an EEG signal is: ; in, The first in the diagram representation Nodes For the Importance score of graph convolutional networks (GCNs) These are weighting coefficients, and they have... , The first in the diagram representation Nodes and PageRank centrality between them The first in the diagram representation Nodes In the Feature vectors extracted by the graph convolutional network (GCN). Adjacency matrix of channel relationships The layer.

[0006] Optionally, in step S102, when combining importance scores and different node masks to mask non-critical nodes layer by layer as input to each level of the graph convolutional network (GCN), any... The functional expression for the node mask of a graph convolutional network (GCN) is: ; in, For the first Node masks for graph convolutional networks (GCNs) The first in the diagram representation Nodes For the Importance score of graph convolutional networks (GCNs) For the first The adaptive threshold of a graph convolutional network (GCN) varies depending on the level of the network. Nodes with a node mask of 0 are considered non-critical nodes. The number of nodes in the graph representation after masking non-key nodes of the input to a graph convolutional network (GCN). for ,in Let be the total number of nodes in the graph representation. The attenuation rate, for Round down, Let G be the number of levels in the graph convolutional network GCN.

[0007] Optionally, in step S102, the function expression for fusing the features extracted by the multi-level graph convolutional network (GCN) across layers to obtain the final extracted fused features is: ; in, The final extracted fusion features, For splicing operations, ~ For the 1st to The linear transformation result of the output features of the graph convolutional network (GCN) of level 1, and any 1st level... Linear transformation results of the output features of the graph convolutional network (GCN) The expression for the computation function is: ; in, For flattened Victor Output features of graph convolutional networks (GCNs) for Dimensional features, For use in 3D feature projection to The transformation matrix of a unified representation space of dimensions. For learnable bias terms, To unify the dimensions of the representation space.

[0008] Optionally, in step S102, the channel relationship adjacency matrix is ​​dynamically updated using the gradient descent method. The function expression is: ; in, This is the updated channel relationship adjacency matrix. This indicates an update operation. To control the learning rate for updating the step size, This is the channel relationship adjacency matrix before the update. This is the gradient update term, which updates the channel relation adjacency matrix. loss function The partial derivative of the channel relationship adjacency matrix with respect to the updated channel relationship adjacency matrix loss function The function expression is: ; in, For cross-entropy loss, and These are the true labels and prediction results for emotion recognition from electroencephalogram (EEG) signals. The regularization coefficient is . Network parameters for a hierarchical dynamic masked image convolutional network The L2 norm.

[0009] Optionally, before step S101, the system may further perform end-to-end training on the hierarchical dynamic mask map convolutional network to update the network parameters of the hierarchical dynamic mask map convolutional network. Furthermore, it updates the network parameters of the hierarchical dynamic mask map convolutional network. The loss function used at that time for: ; ; in, Classify the losses for the task. and These are the weighting coefficients. For the loss of mission return, For the sparsity constraint of graph representation, Adjacency matrix of channel relationships The L1 norm, Adjacency matrix of channel relationships The Middle The electrode channel and the first Functional connections between electrode channels.

[0010] Optionally, when classifying the fused features extracted by the hierarchical dynamic mask image convolutional network using the classification module in step S103 to obtain the final emotion classification result, the method further includes: S201, for the input graph representation The calculation diagram is shown below. any of the first Nodes In the Feature vectors extracted by a graph convolutional network (GCN) The Middle SHAP value of each feature : ; in, for remove All possible sets of nodes other than To represent The set of nodes, for a subset of nodes Let be the total number of nodes in the graph representation. For the classification module, for and The union of; S202, calculate the channel importance of each electrode channel according to the following formula: ; in, For the first The importance of each electrode channel Let GCN be the number of levels in a graph convolutional network. To unify the dimensionality of the representation space, the layer importance of each level of the graph convolutional network (GCN) is calculated according to the following formula: ; in, For the first Layer importance in graph convolutional networks (GCNs).

[0011] Furthermore, the present invention also provides an EEG signal emotion recognition system based on hierarchical dynamic masking, comprising a microprocessor and a memory interconnected thereto, wherein the microprocessor is programmed or configured to execute the EEG signal emotion recognition method based on hierarchical dynamic masking.

[0012] Furthermore, the present invention also provides a computer-readable storage medium storing a computer program or instructions that are programmed or configured to execute the EEG signal emotion recognition method based on hierarchical dynamic masking by a processor.

[0013] In addition, the present invention also provides a computer program product, including a computer program or instructions that are programmed or configured to execute the EEG signal emotion recognition method based on hierarchical dynamic masking via a processor.

[0014] Compared with existing technologies, the present invention mainly achieves the following beneficial effects: The present invention combines the channel relationship adjacency matrix This invention calculates the importance score of each node in the graph representation of EEG signals. By combining the importance score with different node masks, non-critical nodes are masked layer by layer as input to each level of the graph convolutional network (GCN). This achieves selective suppression of high-weighted but spurious connections based on supervised training, while preserving weak but physiologically meaningful synchronization patterns. Through a dynamic node masking mechanism, low-information nodes are gradually masked while key nodes are always retained. This invention can guide the alignment of source and target domain data conditional distributions during the domain adaptation process of multi-source domain adaptation models, enabling cross-subject and cross-session feature analysis and alignment, and improving the accuracy and generalization performance of cross-domain EEG signal recognition. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the basic process of the method in an embodiment of the present invention.

[0016] Figure 2 This is a schematic diagram illustrating the working principle in an embodiment of the present invention. Detailed Implementation

[0017] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings in the embodiments of the present invention.

[0018] like Figure 1 and Figure 2 As shown, the EEG signal emotion recognition method based on hierarchical dynamic masking in this embodiment includes the following steps: S101, A graph structure is constructed for the input EEG signal to generate a graph representation of EEG signals in multiple specified frequency bands. In the graph representation of the EEG signal, nodes represent the EEG signal characteristics of the electrode channels, and edges represent the functional connections between the electrode channels; the channel relationship adjacency matrix is ​​initialized. ; S102, the graphical representations of EEG signals in each specified frequency band are used to extract fusion features using a hierarchical dynamic masked graph convolutional network. This hierarchical dynamic masked graph convolutional network includes a multi-level graph convolutional network (GCN) combined with a channel relation adjacency matrix. The importance score of each node in the graph representation of the EEG signal is calculated. Non-critical nodes are masked layer by layer using the importance score and different node masks as input to each level of the Graph Convolutional Network (GCN). Features extracted by the multi-level GCN are fused across layers to obtain the final extracted fused features. The channel adjacency matrix is ​​dynamically updated using gradient descent. ; S103 uses the classification module to classify the fused features extracted by the hierarchical dynamic mask image convolutional network to obtain the final emotion classification result.

[0019] In step S101, when establishing a graph structure for the input EEG signal to generate a graph representation of EEG signals with multiple specified frequency bands, the multiple specified frequency bands can be specified according to actual needs. For example, as an optional implementation, commonly used EEG signals can be used. , , , and Five frequency bands. The extraction of EEG signal features from each electrode channel includes: S1.1, downsampling the original EEG signal and bandpass filtering from 0.5 to 50 Hz; S1.2, dividing the signal into 1-second non-overlapping time windows; S1.3, calculating the differential entropy features of the five frequency bands within each time window as EEG signal features, i.e., node attributes.

[0020] In step S102, the channel relationship adjacency matrix is ​​combined. The functional expression for calculating the importance score of each node in the graphical representation of an EEG signal is: ; in, The first in the diagram representation Nodes For the Importance score of graph convolutional networks (GCNs) These are weighting coefficients (used to balance structural importance and feature importance), and they have In this embodiment =0.5, The first in the diagram representation Nodes and PageRank centrality between them The first in the diagram representation Nodes In the Feature vectors extracted by the graph convolutional network (GCN). Adjacency matrix of channel relationships The layer.

[0021] As an optional implementation, the node masking mechanism in this embodiment adopts a graph sparsity strategy, increasing the masking ratio layer by layer. Specifically, in step S102, when non-critical nodes are masked layer by layer as input to each level of the graph convolutional network (GCN) by combining importance scores and different node masks, any... The functional expression for the node mask of a graph convolutional network (GCN) is: ; in, For the first Node masks for graph convolutional networks (GCNs) The first in the diagram representation Nodes For the Importance score of graph convolutional networks (GCNs) For the first The adaptive threshold of a graph convolutional network (GCN) varies depending on the level of the network. Nodes with a node mask of 0 are considered non-critical nodes. The number of nodes in the graph representation after masking non-key nodes of the input to a graph convolutional network (GCN). for ,in Let be the total number of nodes in the graph representation. The attenuation rate, for Round down, Let G be the number of levels in the graph convolutional network GCN.

[0022] In step S102 of this embodiment, the hierarchical dynamic mask graph convolutional network includes a multi-level graph convolutional network (GCN). The graph convolutional network (GCN) can be selected as needed, such as a graph convolutional network (GCN) that uses Chebyshev polynomials to implement spectral graph convolution.

[0023] This embodiment is based on hierarchical dynamic masking and graph convolution operations, constructing the final extracted fused features through hierarchical aggregation. The function expression for fusing the features extracted by the multi-level graph convolutional network (GCN) across layers in step S102 to obtain the final extracted fused features is: ; in, The final extracted fusion features, For splicing operations, ~ For the 1st to The linear transformation result of the output features of the graph convolutional network (GCN) of level 1, and any 1st level... Linear transformation results of the output features of the graph convolutional network (GCN) The expression for the computation function is: ; in, For flattened Victor Output features of graph convolutional networks (GCNs) For use in 3D feature projection to The transformation matrix of a unified representation space of dimensions. For learnable bias terms, To unify the dimensions of the representation space. Through the above, the first... The output features of the graph convolutional network (GCN) are linearly transformed to a unified representation space, which ensures the dimensionality compatibility of subsequent feature fusion and achieves cross-layer feature dimension standardization while maintaining the discrimination mode of each layer.

[0024] In step S102, the channel relationship adjacency matrix is ​​dynamically updated using the gradient descent method. The function expression is: ; in, This is the updated channel relationship adjacency matrix. This indicates an update operation. To control the learning rate for updating the step size, This is the channel relationship adjacency matrix before the update. This is the gradient update term, which updates the channel relation adjacency matrix. loss function The partial derivative of the channel relationship adjacency matrix with respect to the updated channel relationship adjacency matrix loss function The function expression is: ; in, For cross-entropy loss, and These are the true labels and prediction results for emotion recognition from electroencephalogram (EEG) signals. The regularization coefficient is . Network parameters for a hierarchical dynamic masked image convolutional network Given the L2 norm, the gradient update term can be expressed as: ; In this context, the first term on the right is the cross-entropy gradient term, and the second term is the regularization gradient term. This is used to scale the eigenvalue matrix for approximation of Chebyshev polynomials.

[0025] Furthermore, prior to step S101 in this embodiment, end-to-end training of the hierarchical dynamic mask map convolutional network is performed to update the network parameters of the hierarchical dynamic mask map convolutional network. Furthermore, it updates the network parameters of the hierarchical dynamic mask map convolutional network. The loss function used at that time for: ; ; in, The task classification loss is used (cross-entropy loss is used in this embodiment). and These are the weighting coefficients. The task regression loss is used (mean squared error loss is used in this embodiment). For the sparsity constraint of graph representation, Adjacency matrix of channel relationships The L1 norm, Adjacency matrix of channel relationships The Middle The electrode channel and the first Functional connections between electrode channels.

[0026] As an optional implementation, this embodiment also provides an interpretability analysis based on SHAP values, which can provide an assessment of the importance of EEG emotion recognition in two dimensions: channel importance and layer importance. In step S103 of this embodiment, when classifying the fusion features extracted by the hierarchical dynamic mask map convolutional network using a classification module to obtain the final emotion classification result, the following is also included: S201, for the input graph representation The calculation diagram is shown below. any of the first Nodes In the Feature vectors extracted by a graph convolutional network (GCN) The Middle SHAP value of each feature : ; in, for remove All possible sets of nodes other than To represent The set of nodes, for a subset of nodes Let be the total number of nodes in the graph representation. For the classification module, for and The union of; S202, calculate the channel importance of each electrode channel according to the following formula (channel importance is used to quantify the global contribution of each EEG electrode channel in all network layers to the final classification decision): ; in, For the first The importance of each electrode channel Let GCN be the number of levels in a graph convolutional network. To unify the dimensionality of the representation space, the layer importance of each level of the Graph Convolutional Network (GCN) is calculated according to the following formula (layer importance evaluates the relative contribution of each GCN layer feature to the overall prediction): ; in, For the first Layer importance in graph convolutional networks (GCNs).

[0027] To verify the EEG signal emotion recognition method based on hierarchical dynamic masking in this embodiment, the SEED dataset, SEED-IV dataset, and DREAMER dataset from the publicly available SEED emotion dataset series were used for cross-subject and cross-conversation emotion recognition analysis. In the SEED dataset, 64 electrodes were placed at standard 10-20 positions to record the EEG signals of 15 test subjects during emotion evoked experiments, with a sampling frequency of 1000Hz. The emotion evoked experiment procedure for the SEED dataset was roughly as follows: A 4-minute edited video clip was played on the experimental screen for each trial, preceded by a 15-second prompt, followed by 45 seconds of self-assessment and a 15-second rest. Each experiment consisted of 15 trials, playing a total of 15 different video stimuli, with 5 clips corresponding to each emotion category. Each subject underwent 3 experiments (one week apart), each consisting of 15 trials, for a total of 45 trials, for the recognition of positive, neutral, and negative emotions. In this way, each participant in each session (the same experiment performed by all participants is considered as one session) takes turns as the target domain, and the remaining 14 participants take turns as the source domain. The average accuracy of the three sessions is used as the recognition result. Accuracy (ACC) and standard deviation (STD) are used as two indicators. Eight state-of-the-art baseline methods for participant emotion recognition, including SVM (Support Vector Machine), GCNN (Graph Convolutional Neural Network), DGCNN (Depth Graph Convolutional Neural Network), BiHDM (Bidirectional Hyperdimensional Memory Network), RGNN (Relational Graph Neural Network), GCB-net (Graph Convolutional-Attention Network), V-IAG (Visual-EEG Integrated Attention Graph Network), and HD-GCN (Hyperdimensional Graph Convolutional Network), are used as comparisons with the method in this embodiment. The final results are shown in Figure 1 and Table 2.

[0028] Table 1: Performance Comparison of Emotion Recognition on SEED and SEED-IV Datasets

[0029] As shown in Table 1, the method in this embodiment performs excellently, achieving classification accuracies of 97.65% (SEED) and 88.38% (SEED-IV), respectively, demonstrating stable advantages across different emotion classification tasks. Among all the comparative methods, the model with the smallest standard deviation indicates stronger robustness across different subject populations.

[0030] Table 2: Comparison of participant dependence experiments on the DREAMER dataset with the strong baseline.

[0031] As shown in Table 2, compared with traditional machine learning methods (Support Vector Machine SVM) and recent graph-based methods (Depth Graph Convolutional Network DGCNN, Graph Convolutional Network GCB-net), the method in this embodiment not only maintains a competitive standard deviation, but also achieves a significantly improved classification accuracy.

[0032] To further verify the performance of the deep learning model of this embodiment on cross-subject and cross-session tasks, it was compared with the current advanced methods for comparison. The results are shown in Table 3.

[0033] Table 3: Performance Comparison of Different Methods

[0034] To systematically evaluate the impact of different mask ratios on emotion recognition performance, the classification accuracy of the first two layers under different mask configurations in this embodiment is shown in Table 4. Table 4: Classification accuracy (%) for different mask ratios on the SEED dataset

[0035] As shown in Tables 3 and 4, the results reveal a clear performance distribution characteristic of the method in this embodiment: the system achieves optimal accuracy when both the first and second layers use a moderate mask ratio. The method in this embodiment outperforms other existing methods in terms of accuracy, and the comparative results demonstrate its effectiveness across subjects and sessions.

[0036] To evaluate the contribution of the loss function to the method in this embodiment, an ablation experiment was designed to study the ablation of conditional entropy loss, mixed loss, and the ablation of sentiment classification results. The ablation experiment analysis in this embodiment is as follows: 1) Ablation of conditional entropy loss during training. The results show that pushing the decision boundary away from the high-density region in the target domain helps improve the model's performance, but the decrease in accuracy is relatively small, about 1%. In addition, even without conditional entropy loss, the results of this method are better than most of the comparison methods mentioned above. 2) The model accuracy decreased significantly after ablation of mixed loss, with a decrease of 2.89% and 2.57% on the SEED dataset. This decrease may be due to the model not considering the conditional distribution between domains, reflecting the necessity and effectiveness of using mixed samples for conditional alignment. 3) When conditional entropy loss and mixed loss are removed, the model is forced to focus only on classification and MMD loss, resulting in a significant performance decrease compared to the complete model. In this embodiment, the frequency domain features of the EEG signal are extracted and a mixed sample set is constructed based on the sample similarity metric. This effectively combines the label information of the source domain samples with the data information of the target domain samples. Furthermore, by optimizing the loss function, the conditional distribution differences between the source and target domains are reduced, resulting in a more accurate recognition effect.

[0037] In summary, the method in this embodiment combines the channel relationship adjacency matrix. This method calculates the importance score of each node in the graph representation of EEG signals. By combining the importance score with different node masks, non-critical nodes are masked layer by layer as input to each level of the graph convolutional network (GCN). This achieves selective suppression of high-weighted but spurious connections based on supervised training, while preserving weak but physiologically meaningful synchronization patterns. Through a dynamic node masking mechanism, low-information nodes are gradually masked while key nodes are always retained. This method can guide the alignment of source and target domain data conditional distributions during the domain adaptation process of multi-source domain adaptation models, enabling cross-subject and cross-session feature analysis and alignment, and improving the accuracy and generalization performance of cross-domain EEG signal recognition.

[0038] Furthermore, this embodiment also provides a brainwave signal emotion recognition system based on hierarchical dynamic masking, including a microprocessor and a memory interconnected, wherein the microprocessor is programmed or configured to execute the brainwave signal emotion recognition method based on hierarchical dynamic masking. Additionally, this embodiment provides a computer-readable storage medium storing a computer program or instructions programmed or configured to execute the brainwave signal emotion recognition method based on hierarchical dynamic masking via a processor. Furthermore, this embodiment also provides a computer program product including a computer program or instructions programmed or configured to execute the brainwave signal emotion recognition method based on hierarchical dynamic masking via a processor.

[0039] Those skilled in the art will understand that the technical solutions provided by this invention may take the form of a method, system, or computer program product. Therefore, this invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this invention may take the form of a computer program product embodied on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, produce an implementation of the flowchart... Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0040] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for emotion recognition of EEG signals based on hierarchical dynamic masking, characterized in that, Includes the following steps: S101, A graph structure is constructed for the input EEG signal to generate a graph representation of EEG signals in multiple specified frequency bands. In the graph representation of the EEG signal, nodes represent the EEG signal characteristics of the electrode channels, and edges represent the functional connections between the electrode channels; the channel relationship adjacency matrix is ​​initialized. ; S102, the graphical representations of EEG signals in each specified frequency band are used to extract fusion features using a hierarchical dynamic masked graph convolutional network. This hierarchical dynamic masked graph convolutional network includes a multi-level graph convolutional network (GCN) combined with a channel relation adjacency matrix. The importance score of each node in the graph representation of the EEG signal is calculated. Non-critical nodes are then masked layer by layer using the importance score and different node masks as input to each level of the Graph Convolutional Network (GCN). Features extracted by the multi-level GCN are fused across layers to obtain the final extracted fused features. The channel adjacency matrix is ​​dynamically updated using gradient descent. ; S103 uses the classification module to classify the fused features extracted by the hierarchical dynamic mask image convolutional network to obtain the final emotion classification result.

2. The EEG signal emotion recognition method based on hierarchical dynamic masking according to claim 1, characterized in that, In step S102, the channel relationship adjacency matrix is ​​combined. The functional expression for calculating the importance score of each node in the graphical representation of an EEG signal is: ； in, The first in the diagram representation Nodes For the Importance score of graph convolutional networks (GCNs) These are weighting coefficients, and they have... , The first in the diagram representation Nodes and PageRank centrality between them The first in the diagram representation Nodes In the Feature vectors extracted by the graph convolutional network (GCN). Adjacency matrix of channel relationships The layer.

3. The EEG signal emotion recognition method based on hierarchical dynamic masking according to claim 1, characterized in that, In step S102, when combining importance scores and different node masks to mask non-critical nodes layer by layer as input to each level of the graph convolutional network (GCN), any... The functional expression for the node mask of a graph convolutional network (GCN) is: ； in, For the first Node masks for graph convolutional networks (GCNs) The first in the diagram representation Nodes For the Importance score of graph convolutional networks (GCNs) For the first The adaptive threshold of a graph convolutional network (GCN) varies depending on the level of the network. Nodes with a node mask of 0 are considered non-critical nodes. The number of nodes in the graph representation after masking non-key nodes of the input to a graph convolutional network (GCN). for ,in Let be the total number of nodes in the graph representation. The attenuation rate, for Round down to the nearest whole number, Let G be the number of levels in the graph convolutional network GCN.

4. The EEG signal emotion recognition method based on hierarchical dynamic masking according to claim 1, characterized in that, In step S102, the function expression for fusing the features extracted by the multi-level graph convolutional network (GCN) across layers to obtain the final extracted fused features is as follows: ； in, The final extracted fusion features, For splicing operations, ~ For the 1st to The linear transformation result of the output features of the graph convolutional network (GCN) of level 1, and any 1st level... Linear transformation results of the output features of the graph convolutional network (GCN) The expression for the computation function is: ； in, For flattened Victor Output features of graph convolutional networks (GCNs) for dimensional features, For use in 3D feature projection to The transformation matrix of a unified representation space of dimensions. For learnable bias terms, To unify the dimensions of the representation space.

5. The EEG signal emotion recognition method based on hierarchical dynamic masking according to claim 1, characterized in that, In step S102, the channel relationship adjacency matrix is ​​dynamically updated using the gradient descent method. The function expression is: ； in, This is the updated channel relationship adjacency matrix. This indicates an update operation. To control the learning rate for updating the step size, This is the channel relationship adjacency matrix before the update. This is the gradient update term, which updates the channel relation adjacency matrix. loss function The partial derivative of the channel relationship adjacency matrix with respect to the updated channel relationship adjacency matrix loss function The function expression is: ； in, For cross-entropy loss, and These are the true labels and prediction results for emotion recognition from electroencephalogram (EEG) signals. The regularization coefficient is . Network parameters for a hierarchical dynamic masked image convolutional network The L2 norm.

6. The EEG signal emotion recognition method based on hierarchical dynamic masking according to claim 1, characterized in that, Before step S101, the process also includes end-to-end training of the hierarchical dynamic mask map convolutional network to update its network parameters. Furthermore, it updates the network parameters of the hierarchical dynamic mask map convolutional network. The loss function used at that time for: ； ； in, Classify the losses for the task. and These are the weighting coefficients. For the loss of mission return, For the sparsity constraint of graph representation, Adjacency matrix of channel relationships The L1 norm, Adjacency matrix of channel relationships The Middle The electrode channel and the first Functional connections between electrode channels.

7. The EEG signal emotion recognition method based on hierarchical dynamic masking according to claim 1, characterized in that, In step S103, when the fused features extracted by the hierarchical dynamic mask image convolutional network are classified using the classification module to obtain the final emotion classification result, the following is also included: S201, for the input graph representation The calculation diagram is shown below. any of the first Nodes In the Feature vectors extracted by a graph convolutional network (GCN) The Middle SHAP value of each feature : ； in, for remove All possible sets of nodes other than To represent The set of nodes, for a subset of nodes Let be the total number of nodes in the graph representation. For the classification module, for and The union of; S202, calculate the channel importance of each electrode channel according to the following formula: ； in, For the first The importance of each electrode channel Let GCN be the number of levels in a graph convolutional network. To unify the dimensionality of the representation space, the layer importance of each level of the graph convolutional network (GCN) is calculated according to the following formula: ； in, For the first Layer importance in graph convolutional networks (GCNs).

8. A brainwave signal emotion recognition system based on hierarchical dynamic masking, comprising a microprocessor and a memory interconnected, characterized in that, The microprocessor is programmed or configured to perform the EEG signal emotion recognition method based on hierarchical dynamic masking as described in any one of claims 1 to 7.

9. A computer-readable storage medium storing a computer program or instructions, characterized in that, The computer program or instructions are programmed or configured to execute the EEG signal emotion recognition method based on hierarchical dynamic masking as described in any one of claims 1 to 7 via a processor.

10. A computer program product, comprising a computer program or instructions, characterized in that, The computer program or instructions are programmed or configured to execute the EEG signal emotion recognition method based on hierarchical dynamic masking as described in any one of claims 1 to 7 via a processor.

Images