Personalized electroencephalogram emotion recognition method and system based on dynamic brain region division

Abstract

The invention discloses a personalized electroencephalogram emotion recognition method and system based on dynamic brain region division, and belongs to the technical field of electroencephalogram emotion recognition. The method comprises the steps that electroencephalogram signals are preprocessed, and an electroencephalogram signal matrix is constructed; on the basis of the preprocessed electroencephalogram signals, five-frequency-band differential entropy feature values are extracted, multi-head self-attention calculation is carried out in combination with rotation position coding, and channel feature representation with strong emotion intensity is output; on the basis of the constructed electroencephalogram signal matrix, dynamically dividing a brain function region by using a CNM algorithm, optimizing the structure of the brain function region by using a modularity value, and outputting brain region characteristics; and fusing the channel feature representation and the brain region features, predicting an emotion tag through a graph neural network, and outputting an emotion recognition result. According to the method, the individual brain function region can be dynamically divided, the function integration information of the brain region is effectively captured, and deep fusion of brain region function connection and channel emotion information is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of EEG emotion recognition, and specifically relates to a personalized EEG emotion recognition method and system based on dynamic brain region division. Background Art

[0002] The statements herein merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Human emotional states are closely related to physical health and behavioral patterns. Physiological signals can better express true emotions than behavioral signals. Among many physiological signals, electroencephalography (EEG), as a non-invasive signal, can directly measure brain electrical activity during emotional states. It offers advantages such as high temporal resolution, high transmission speed, low cost, and portability. Consequently, research on EEG-based emotion recognition has become increasingly popular and is a key research topic in affective computing, human-computer interaction, and brain-computer interfaces. Existing work largely utilizes deep learning methods such as convolution, graph convolution, and attention mechanisms to model temporal and channel-to-channel dependencies of EEG signals based on known time-frequency domain features. For example, Yi et al. exploited the high temporal resolution and asymmetric spatial activation characteristics of EEG to identify emotions. They performed convolution in the temporal and channel dimensions to extract temporal and spatial features, respectively, and then fused these features for emotion classification. Zhang et al. utilized a hierarchical self-attention network to capture local and global temporal features of emotion and modeled spatial information using a channel-wise correlation loss. Salanitri et al. embedded the differential entropy features of EEG signals into the electroencephalogram to express spatial information, and used multi-scale convolutional networks and TimesNet networks to extract cross-channel spatial and temporal features. In addition, some methods consider that the key to various brain functions lies in the connection and communication between different regions, and introduce brain region connectivity. For example, Yi et al. divided the scalp area into 17 functional areas. The signal of each area is represented by a regional token, which interacts with the original channel nodes to form a multi-level hierarchical structure. Ding et al. constructed a local-global graph network. Each local graph represents a specific functional area of ​​the brain, and the global connection between local graphs reflects the complex functional connection relationship between different functional areas of the brain.

[0004] In reality, the brain's neural networks are dynamic. Not only do the size and location of brain regions vary significantly between individuals, but different people can exhibit distinct patterns of brain activity when performing the same task. For the same individual, the brain also dynamically reconfigures its functional organization across different cognitive tasks. Existing work uses fixed brain region parcellations based on resting-state measurements of brain region structure across multiple subjects and the average of these parcellations. Gordon et al. compared the similarity between group-averaged and individual-level brain region connectivity and analyzed how this similarity varied across brain regions and subjects. They found that group-individual similarity in connectivity patterns increased with time points, with most remaining below 0.6, and individual similarities reaching less than 0.35. When the number of time points exceeded 300 (equivalent to 12.5 minutes), the similarity between subjects' brain regions was higher, reaching a maximum of less than 0.7. However, most current methods use training samples of 1-3 seconds, roughly equivalent to one time point. Even using unsliced ​​data, the sample length is only about four minutes. Therefore, the group-averaged and individual-level brain region distributions used in existing studies differ significantly. This group template cannot match the emotion datasets used in existing studies, nor can it correctly model the distribution of brain areas in different individuals. Summary of the Invention

[0005] The purpose of the present invention is to overcome the deficiencies in the above-mentioned prior art and to provide a personalized EEG emotion recognition method and system based on dynamic brain area division, which can dynamically divide individual brain functional areas, effectively capture the functional integration information of brain areas, and realize the deep integration of brain area functional connections and channel emotional information.

[0006] In order to achieve the above object, the present invention is implemented through the following technical solutions: On the one hand, the technical solution of the present invention provides a personalized EEG emotion recognition method based on dynamic brain region division, including: Preprocess the EEG signal and construct the EEG signal matrix; Based on the preprocessed EEG signal, the five-band differential entropy eigenvalues ​​are extracted and combined with rotational position encoding to perform multi-head self-attention calculation to output channel feature representations with strong emotional intensity. Based on the constructed EEG signal matrix, the CNM algorithm is used to dynamically divide brain functional areas, and the modularity value is used to optimize the brain functional area structure and output brain area characteristics; By integrating channel feature representation and brain region features, the emotion label is predicted through graph neural network and the emotion recognition result is output.

[0007] In at least one embodiment, preprocessing the EEG signal includes removing the baseline signal and signal slicing, specifically: calculating the average value of the baseline signal; subtracting the average value of the baseline signal from the EEG signal to obtain the EEG signal after removing the baseline signal; and dividing the EEG signal after removing the baseline signal into several slices, each slice being an EEG sample.

[0008] In at least one embodiment, constructing the EEG signal matrix is ​​specifically as follows: for each EEG signal, calculating the weighted phase lag index between all channels in pairs, thereby constructing a functional connectivity matrix reflecting the brain functional network, and retaining the top 20% of the connections in terms of strength.

[0009] In at least one embodiment, based on the emotion intensity attention module, the preprocessed EEG signal is decomposed into five sub-bands, and the differential entropy eigenvalue is extracted from each sub-band as a feature representation; the rotation position code is embedded in each feature representation and multi-head self-attention calculation is performed; the output of each self-attention head is spliced ​​as the final output of the emotion intensity attention module, and a channel feature representation with strong emotion intensity is output.

[0010] In at least one embodiment, the emotion intensity attention module uses margin-based negative sample loss as the loss function, which is expressed as: ; Where, is the distance between samples; The margin is the distance between samples. is the sample number; is the total number of samples; is the final extracted emotional feature representation.

[0011] In at least one embodiment, modularity and modularity increment matrix elements for: ; ; ; in, Indicates sub-area The ratio of internal edges to all edges in the graph, Indicates sub-area With sub-area The ratio of the number of connected edges to the total number of edges ; Indicates connection to a sub-area The ratio of edges in to the total number of edges; For sub-area With sub-area The increase in modularity after merging; sub-areas and When there are edges connecting them, It is 1 when it is not connected and 0 when it is not connected. Initially, it is assumed that each node is an independent sub-area, and sub-areas are continuously merged until all nodes in the network are included in one area, or the maximum modularity increment is negative.

[0012] In at least one embodiment, the overall feature of the brain function region is represented and output as a weighted sum of all nodes within the brain function region. The overall feature of the brain function region is represented as: ; Where, is the Hadamard product; is the number of channels in a certain brain functional area; It is a channel feature representation with strong emotional intensity; is the degree centrality weight, which assigns a value to each node to indicate the degree of connection between the node and other nodes in the network.

[0013] In at least one embodiment, the emotion prediction module uses cross entropy loss as the loss function, which is expressed as: ; Where, is the true label vector, is the probability distribution of the model’s predicted sentiment labels; is the number of samples in a batch; express norm; Represents the trainable parameters of the model; the margin-based negative sample loss Together with the categorical cross entropy loss, it serves as the total loss function during model training.

[0014] In at least one embodiment, the margin-based negative sample loss and the categorical cross entropy loss are combined. Together they serve as the total loss function during model training, expressed as: ; Where, is the hyperparameter, that is, the weight of the margin-based negative sample loss.

[0015] On the other hand, the technical solution of the present invention also provides a personalized EEG emotion recognition system based on dynamic brain region division, including: The data preprocessing module is configured to: preprocess the EEG signal and construct an EEG signal matrix; The emotion intensity attention module is configured to: extract the five-band differential entropy eigenvalues ​​based on the preprocessed EEG signal, perform multi-head self-attention calculations in combination with rotational position encoding, and output channel feature representations with strong emotion intensity; The functional space partitioning module is configured to: dynamically divide brain functional areas using the CNM algorithm based on the constructed EEG signal matrix, optimize the brain functional area structure with modularity value, and output brain area features; The emotion prediction module is configured to: fuse channel feature representation and brain region features, predict emotion labels through graph neural networks, and output emotion recognition results.

[0016] The beneficial effects of the technical solution of the present invention are as follows: The personalized EEG emotion recognition method based on dynamic brain area division of the present invention adaptively divides personalized brain functional areas through the functional space division module, reduces the distribution differences between subjects, extracts key emotional features from the EEG signal through the emotion intensity attention module, suppresses atypical noise, and finally uses the graph neural network architecture to fuse the outputs of the emotion intensity attention module and the functional space division module to model the dynamic functional connections between brain functional areas, effectively capture the functional integration information of brain areas, and realize the deep fusion of brain area functional connections and channel emotional information. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0018] Figure 1 This is a schematic diagram of the overall framework of a personalized EEG emotion recognition network based on dynamic brain area division in Example 1 of the present invention. DETAILED DESCRIPTION

[0019] It should be noted that the following detailed description is illustrative and is intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0020] As introduced in the background technology, the purpose of the present invention is to overcome the shortcomings of the above-mentioned existing technologies and provide a personalized EEG emotion recognition method and system based on dynamic brain area division, which can dynamically divide individual brain functional areas, effectively capture the functional integration information of brain areas, and realize the deep integration of brain area functional connections and channel emotional information.

[0021] Example 1 In a typical embodiment of the present invention, in order to mine more available and competitive EEG emotion features from brain area activities that are more in line with individual characteristics and emotional task requirements, this embodiment proposes a personalized EEG emotion recognition network based on dynamic brain area division (PEER-DER), the overall framework of which is as follows: Figure 1 As shown, it consists of four important modules, namely data preprocessing module, emotion intensity attention module, functional space division module and emotion prediction module. Based on this, this embodiment discloses a personalized EEG emotion recognition method based on dynamic brain region division, including: S1, preprocessing EEG signals and constructing functional connectivity matrix; S2. Extract the five-band differential entropy eigenvalues ​​based on the preprocessed EEG signal, perform multi-head self-attention calculations in combination with rotational position encoding, and output channel feature representations with strong emotional intensity. S3. Based on the constructed functional connectivity matrix, the CNM algorithm is used to dynamically divide the brain functional areas, the modularity value is used to optimize the brain functional area structure, and the brain area characteristics are output; S4. Fusion channel feature representation and brain region features, predict emotion labels through graph neural network, and output emotion recognition results.

[0022] The personalized EEG emotion recognition method based on dynamic brain area division is described in detail below in conjunction with specific implementation methods.

[0023] S1. Preprocess the EEG signal and construct the functional connectivity matrix.

[0024] like Figure 1 As shown in part (a), the data preprocessing module of the PEER-DBR network of this embodiment includes two functions: one is to preprocess the EEG signal, and the other is to construct a functional connectivity matrix using the EEG signal, as described below.

[0025] (1) Preprocessing of EEG signals.

[0026] In public databases related to EEG emotion recognition (such as DEAP, SEED, SEED-IV), EEG signal data usually contains baseline signals and task-induced test signals. In view of this feature, the data preprocessing module of this embodiment uses the baseline signal removal and sliding window technology to process the acquired raw EEG signals, which can effectively improve the performance of the PEER-DBR network in emotion recognition. Specifically, Hz sampling frequency, The EEG signal data collected at the sampling time is expressed as ,in, is the number of electrode nodes, is the number of sampling points, Indicates the number of sampling points The baseline signal, Indicates the number of sampling points The average value of the baseline signal is It can be expressed as: (1); Then, the average value of the baseline signal is subtracted from each trial signal to remove the baseline drift, and the EEG signal after removing the baseline signal is obtained, which can be expressed as: (2); In order to increase the amount of training data, each EEG signal after removing the baseline signal Split into several slices ,in Indicates the i EEG slices, Indicates the number of sampling points in each sliding window.

[0027] (2) Use EEG signals to construct a functional connectivity matrix.

[0028] In this embodiment, the data preprocessing module uses a phase synchronization strategy to construct a functional connectivity matrix. How the various brain regions work together in a specific task or state can be reflected by the synchronization phenomenon, an indicator reflecting the information exchange between different brain regions. On this basis, the phase synchronization strategy generally refers to a method of evaluating the relationship between their phases when analyzing two or more signals. In neuroscience, the phase synchronization strategy has been widely used in the analysis of functional connectivity between different brain regions. This method can reveal the phase relationship between brain regions at a specific frequency, thereby helping to understand the dynamic changes of the brain's functional network. There are many existing phase synchronization methods. In order to reduce the influence of volume conduction and noise, the data preprocessing module of this embodiment adopts the weighted phase lag index (WPLI), which is expressed as follows: (3); The phase synchronization index is the imaginary part of the cross-spectrum between the signals, Indicates whether the signal from the first sensor tends to lead or lag the signal from the second sensor in phase. The strength of the phase synchronization can be measured by The size of The WPLI is an improvement on the phase lag index (PLI), which not only retains the advantage of reducing the influence of volume conduction effect, but also enhances the sensitivity to phase changes with smaller intensity but practical physiological significance.

[0029] In all data sets, each trial of each participant was treated as an independent sample. For each sample, the pairwise weighted phase lag index between all channels (i.e., all monitoring electrodes) was calculated, and a functional connectivity matrix reflecting the brain functional network was constructed based on this.

[0030] Brain functional networks contain numerous weak and spurious connections, which often obscure the topological structure of core connections. To address this issue, this example avoids using a threshold-based binarization method. Instead, it retains only the 20% strongest connections, resulting in a weighted graph topology that preserves more biologically relevant information.

[0031] S2. Based on the preprocessed EEG signal, the five-band differential entropy eigenvalues ​​are extracted, and multi-head self-attention calculation is performed in combination with rotational position encoding to output channel feature representation with strong emotional intensity.

[0032] Because emotional states are unstable and discrete over long periods of time, expected strong emotions typically only last for a few seconds. In other words, EEG data may have a low signal-to-noise ratio at the temporal level. Therefore, the PEER-DBR network of this embodiment designs an emotion intensity attention module to focus on key information in emotional states. It extracts key emotional features from EEG signals based on contrastive learning and suppresses atypical noise, thereby improving the performance of emotion recognition. Figure 1 As shown in part (b).

[0033] Clinically, there are five main rhythms of EEG signals: Wave (8-13Hz), Wave (13-30Hz), Wave (4-8Hz), wave (≤3Hz) and Wave (≥30Hz), therefore, in this embodiment, the discrete Fourier transform (DFT) method is first used to decompose the pre-processed EEG signal according to the above rhythm to obtain five sub-bands , and extract the differential entropy (DE) eigenvalue on each sub-band as the feature representation of the sample .

[0034] Although convolution and recurrence operations are suitable for extracting EEG features, they are both local, that is, they only act on several neighbors of the target element. Taking into account the fragmentary nature of emotional intensity, this embodiment uses global attention to implement the emotional intensity attention module, focusing on fragments with stronger emotional states, and extracting EEG features that contain more emotional information. In order to improve the model's sensitivity to element position information, the rotational position encoding (RoPE) is introduced and embedded in each feature representation. This encoding method takes into account both the absolute position and the relative position of the element, thereby maintaining the flexibility and invariance of the relative position encoding while also utilizing the accuracy and globality of the absolute position encoding. Represent each feature The rotation position encoding is embedded in the model and multi-head self-attention calculation is performed. The output of each self-attention head is concatenated as the final output of the emotion intensity attention module, and the channel feature representation with strong emotion intensity is output.

[0035] Specifically, first, the feature representation of the embedded rotation position encoding is converted into a query vector , key vector Sum value vector : (4); Where, Representing the overall features, i.e., the feature representation embedded in the rotation position encoding; Represents the key vector Time The transformation matrix corresponding to the attention head; Indicates the The query vector for each attention head.

[0036] Each head computes its self-attention using dot-product scaling, i.e. , then the self-attention is expressed as: (5); The output of each self-attention head is concatenated as the final output of the emotion intensity attention module to obtain a channel feature representation with strong emotion intensity, which can be expressed as: (6); Where, represents the final extracted emotional feature representation, Indicates the Feature representation obtained by the attention head; Represents the output transformation matrix.

[0037] Following the self-attention layer are layer normalization, GELU, linear layers, and dropout operations, with residual connections employed to prevent network degradation. Layer normalization adjusts the input data distribution, improving model training speed and stability; the GELU activation function introduces nonlinearity; the linear layer maps features to the output space, helping the model learn complex feature representations; and dropout prevents overfitting.

[0038] In order to enhance the accuracy of the channel feature representation extracted by the emotion intensity attention module, this embodiment uses margin-based negative sample loss as the loss function, which is expressed as: (7); Where, is the distance between samples; The margin is the distance between samples. is the sample number; is the total number of samples.

[0039] Set a minimum distance on the negative sample pairs, forcing the distance between the negative sample pairs to be greater than , if the distance is less than , a loss will be generated, which encourages the model to increase the distance between negative sample pairs, that is, to make the channel feature representations belonging to different emotion labels stay away from each other, while there is no restriction on the channel features belonging to the same label.

[0040] S3. Based on the constructed functional connectivity matrix, the CNM algorithm is used to dynamically divide the brain functional areas, the modularity value is used to optimize the brain functional area structure, and the brain area characteristics are output.

[0041] Due to the volume conduction effect in the brain, the signal of each channel in the collected EEG is the result of the superposition of signals from various brain regions, and there is a lot of noise. The anatomical structure is the precise structure obtained by inserting electrodes into the cerebral cortex. The functional connectivity characteristics obtained using this highly redundant EEG signal cannot accurately reflect the anatomical brain regions and the activities between brain regions. Moreover, the pattern of EEG activity will be affected by factors such as the subject's physiological differences, neurophysiological state, scalp and skull structure, life experience and training, including but not limited to age, gender, fatigue state, scalp thickness, and skull structure. The individual differences in EEG signals limit the generalization ability of the model. Therefore, this embodiment adopts a dynamic method to divide the brain functional areas that are more in line with the subject's own conditions to improve the generalization ability of the model.

[0042] Brain connectivity refers to the synchronous activity of neurons in different brain regions, and functional connectivity is based on the functional integration of brain regions, forming an undirected complex network topology. The brain functional network includes small-world network characteristics and modular structure, that is, the nodes within each module are closely connected, while the connections between different modules are relatively few. Based on this characteristic, the functional space partitioning module of this embodiment uses the CNM algorithm to dynamically divide brain functional areas based on the functional connectivity matrix constructed by the data preprocessing module, and divides each brain functional network into multiple areas with closer internal node connections. Modularity As an indicator of the structural strength of brain functional areas, the larger the value, the better the brain area division effect. Based on this, after dynamically dividing the brain functional areas in this embodiment, the modularity is used to The value optimizes the brain functional regional structure.

[0043] Defining modularity and modularity increment matrix elements for: (8); (9); (10); in, Indicates sub-area The ratio of internal edges to all edges in the graph, Indicates sub-area With sub-area The ratio of the number of connected edges to the total number of edges ; Indicates connection to a sub-area The ratio of edges in to the total number of edges; For sub-area With sub-area The increase in modularity after merging; sub-areas and When there are edges connecting them, When the network is connected, it is 1, and when it is disconnected, it is 0. Initially, it is assumed that each node is an independent sub-region, and sub-regions are continuously merged until all nodes in the network are included in one region, or the maximum modularity increment is negative.

[0044] In order to ensure the completeness of brain region division, a few regions containing only 1 or 2 nodes were merged into a new region. Subsequently, the nodes within the brain functional region were filtered and aggregated, and the weighted sum of all nodes within the brain functional region was used to represent the overall characteristics of the brain functional region. : (11); in, is the Hadamard product; is the number of channels in a certain brain functional area; It is a channel feature representation with strong emotional intensity; is the degree centrality weight. Degree centrality can assign a value to each node to indicate the degree of connection between the node and other nodes in the network.

[0045] S4. Fusion channel feature representation and brain region features, predict emotion labels through graph neural network, and output emotion recognition results.

[0046] In this embodiment, the GCN graph neural network integrates the channel features output by the emotion intensity attention module and the brain region features output by the functional space partitioning module, and predicts and outputs the emotion label through the linear layer and softmax function. Specifically, GCN is first used to model the brain region, and the relationship between them is captured through the information of adjacent brain regions. The brain region activity is constructed as an undirected graph ,in represents a vertex set, Denote the edge set. Let for The normalized Laplace matrix of : (12); in is a diagonal matrix and , represents the adjacency matrix, is the number of brain regions, i.e. In order to capture the similarity between node representations, similarity is used to represent the connection weight between two brain regions, and self-loops are added to the adjacency matrix. The GCN iterations from different layers can be expressed as: (13); in, Indicates the The overall characteristics of the brain functional area of ​​the layer input, is the weight matrix. The final output of the graph convolution predicts the emotion label through the linear layer and the softmax function and outputs the emotion recognition result. At the same time, the classification task of the emotion prediction module uses the cross entropy loss as the loss function, which is expressed as: (14); Where, is the true label vector, is the probability distribution of the model’s predicted sentiment labels; is the number of samples in a batch; express norm; Represents the trainable parameters of the model; the margin-based negative sample loss Together with the categorical cross entropy loss, it serves as the total loss function during model training.

[0047] is the number of emotion labels; is the true label vector; Is the probability distribution of the model's predicted emotional labels. Margin-based negative sample loss and classification cross entropy loss Together they serve as the total loss function during model training, expressed as: (15); Where, is the hyperparameter, that is, the weight of the margin-based negative sample loss.

[0048] In this example, the performance of this method was evaluated using three commonly used EEG emotion datasets: the SEED dataset, the SEED-IV dataset, and the DEAP dataset. Considering that the brain activity patterns corresponding to a subject's own emotional response are stable, while those corresponding to emotional responses between different subjects may differ significantly, for different experiments on the same subject, the average number of brain functional area divisions across multiple experiments was used as the number of brain areas for that subject. Given that EEG signals depend on the subject, the EEG-based emotion recognition task was treated as a subject-dependent problem, and classifiers were trained separately for all participants. The model was implemented using the PyTorch framework and trained on an A100 GPU. Furthermore, the SGD optimizer was used to minimize the total loss function. During training, each sample was 2 seconds long, the batch size was 64, the dimensionality of all hidden vectors was set to 128, and the dimensionality was set to 0.5. When training the network on the DEAP dataset, label smoothing with a smoothing rate of 0.1 was used due to the high class imbalance of some subjects in some DEAP datasets.

[0049] For the DEAP dataset, the following representative models (baseline models) were selected for comparison: Priyasad et al. employed a channel-specific encoder network, geometric deep learning, and multi-channel feature fusion to identify emotional states from EEG data. Zhang et al. proposed a hierarchical self-attention network with a channel-correlation loss, designed to capture local and global temporal patterns and spatial emotional features, thereby identifying EEG segments with high emotional relevance. Zhu et al. proposed a multi-domain feature fusion framework that combines time-frequency and spatial representations via a cross-domain attention mechanism and a two-stage feature integration strategy. Ding et al. introduced TSception, a multi-scale convolutional neural network that effectively captures temporal dynamics and spatial asymmetries in EEG signals. Han et al. proposed MS-ERM, a hybrid model combining multi-scale cellular neural networks with temporal networks to extract time-frequency and spatial features. Wang et al. proposed ALGGNet, which employs a biclustering-based adaptive local region partitioning module to generate dataset-specific functional brain regions and extracts spatiotemporal features through hierarchical graph learning. The results of comparing this method with the aforementioned baseline models on the DEAP dataset are shown in Table 1.

[0050] Table 1

[0051] For the SEED dataset, the following representative models (baseline models) were selected for comparison: Jiang et al. introduced ECO-FET, a contrastive learning framework that fuses EEG functional connectivity features across spectral, spatial, and temporal dimensions with eye movement data to improve emotion recognition while reducing reliance on labeled data; Liu et al. proposed DA-CapsNet, a multi-branch capsule network that employs adversarial domain adaptation to enhance cross-subject emotion recognition performance; Ding et al. proposed EmT, a transformer model that uses a long sliding window to segment each trial with a short stride and further divides long segments into subsegments to mitigate noisy labels and capture both short-term and long-term emotional context; and Tan et al. proposed EEG, a semi-supervised, domain-adaptive framework that leverages adversarial learning and dynamic EEG features for robust cross-subject emotion classification. The method of Zhang et al. has been introduced above and will not be repeated here. The results of comparing our method with the aforementioned baseline models on the SEED dataset are shown in Table 2.

[0052] Table 2

[0053] For evaluation on the SEED-IV dataset, the following representative models (baseline models) were selected for comparison: PR-PL proposed by Zhou et al., a pairwise learning method based on prototype representations designed to mitigate individual differences and label noise; PCDG introduced by Cai et al., a two-stage prototype contrastive domain generalization framework that combines a new CNN architecture (combining residual and CBAM blocks) with contrastive learning to address the challenge of cross-subject recognition; DMMR was developed by Wang et al., a method that applies two-stage training (self-supervised pre-training and supervised fine-tuning) to improve cross-subject generalization capabilities through hidden layer data mixing and noise injection; the methods of Jiang et al. and Tan et al. have been introduced above and will not be repeated here. The results of the comparison of this method with the above baseline models on the SEED-IV dataset are shown in Table 3.

[0054] Table 3

[0055] As can be seen, our method achieved recognition accuracy of 92.18% (average of valence and arousal accuracy), 87.79%, and 78.26% on the DEAP, SEED, and SEED-IV datasets, respectively, surpassing the best recognition results of all listed baseline methods by 1.3%, 1.4%, and 0.46%. This significant improvement in recognition results demonstrates that our framework is able to better capture the underlying emotional characteristics in EEG signals. Functional connectivity features and interactions between functional brain regions facilitate the identification of emotional signals in EEG data. On the DEAP dataset, the average recognition accuracy for the valence and arousal evaluation metrics was 91.51% and 92.86%, respectively. The close accuracy between these two metrics indicates that our model is more stable.

[0056] Example 2 In a typical embodiment of the present invention, this embodiment discloses a personalized EEG emotion recognition system based on dynamic brain region division, including: The data preprocessing module is configured to: preprocess the EEG signal and construct a functional connectivity matrix; The emotion intensity attention module is configured to: extract the five-band differential entropy eigenvalues ​​based on the preprocessed EEG signal, perform multi-head self-attention calculations in combination with rotational position encoding, and output channel feature representations with strong emotion intensity; The functional space partitioning module is configured to: dynamically divide brain functional areas based on the constructed functional connectivity matrix using the CNM algorithm, optimize the brain functional area structure with modularity values, and output brain area features; The emotion prediction module is configured to: fuse channel feature representation and brain region features, predict emotion labels through graph neural networks, and output emotion recognition results.

[0057] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A personalized EEG emotion recognition method based on dynamic brain region division, characterized by: include: Preprocess the EEG signal and construct the EEG signal matrix; Based on the preprocessed EEG signal, the five-band differential entropy eigenvalues ​​are extracted and combined with rotational position encoding to perform multi-head self-attention calculation to output channel feature representations with strong emotional intensity. Based on the constructed EEG signal matrix, the CNM algorithm is used to dynamically divide brain functional areas, and the modularity value is used to optimize the brain functional area structure and output brain area characteristics; By integrating channel feature representation and brain region features, the emotion label is predicted through graph neural network and the emotion recognition result is output.

2. The personalized EEG emotion recognition method based on dynamic brain region division according to claim 1, characterized in that: The preprocessing of the EEG signal includes removing the baseline signal and signal slicing, specifically: calculating the average value of the baseline signal; subtracting the average value of the baseline signal from the EEG signal to obtain the EEG signal after removing the baseline signal; and dividing the EEG signal after removing the baseline signal into several slices, each slice being an EEG sample.

3. The personalized EEG emotion recognition method based on dynamic brain region division according to claim 1, characterized in that: The EEG signal matrix was constructed as follows: for each EEG signal, the weighted phase lag index between all channels was calculated to construct a functional connectivity matrix reflecting the brain functional network, and the top 20% connections were retained.

4. The personalized EEG emotion recognition method based on dynamic brain region division according to claim 1, characterized in that: Based on the emotion intensity attention module, the preprocessed EEG signal is decomposed into five sub-bands, and the differential entropy eigenvalue is extracted from each sub-band as the feature representation; A rotational position code is embedded in each feature representation and multi-head self-attention calculation is performed; the output of each self-attention head is concatenated as the final output of the emotion intensity attention module, which outputs a channel feature representation with strong emotion intensity.

5. The personalized EEG emotion recognition method based on dynamic brain region division according to claim 4, characterized in that: The emotion intensity attention module uses margin-based negative sample loss as the loss function, which is expressed as: ； Where, is the distance between samples; The margin is the distance between samples. is the sample number; is the total number of samples; is the final extracted emotional feature representation.

6. The personalized EEG emotion recognition method based on dynamic brain region division according to claim 1, characterized in that: Defining modularity and modularity increment matrix elements for: ； ； ； in, Indicates sub-area The ratio of internal edges to all edges in the graph, Indicates sub-area With sub-area The ratio of the number of connected edges to the total number of edges ; Indicates connection to a sub-area The ratio of edges in to the total number of edges; For sub-area With sub-area The increase in modularity after merging; sub-areas and When there are edges connecting them, It is 1 when it is not connected and 0 when it is not connected. Initially, it is assumed that each node is an independent sub-area, and sub-areas are continuously merged until all nodes in the network are included in one area, or the maximum modularity increment is negative.

7. The personalized EEG emotion recognition method based on dynamic brain region division according to claim 6, characterized in that: The weighted sum of all nodes in the brain functional area is used to represent the overall characteristics of the brain functional area and output. The overall characteristics of the brain functional area are expressed as: ； Where, is the Hadamard product; is the number of channels in a certain brain functional area; It is a channel feature representation with strong emotional intensity; is the degree centrality weight, which assigns a value to each node to indicate the degree of connection between the node and other nodes in the network.

8. The personalized EEG emotion recognition method based on dynamic brain region division according to claim 1, characterized in that: The emotion prediction module uses cross entropy loss as the loss function, which is expressed as: ； Where, is the true label vector, is the probability distribution of the model’s predicted sentiment labels; is the number of samples in a batch; express norm; Represents the trainable parameters of the model; the margin-based negative sample loss Together with the categorical cross entropy loss, it serves as the total loss function during model training.

9. The personalized EEG emotion recognition method based on dynamic brain region division according to claim 7, characterized in that: Combine margin-based negative sample loss and categorical cross entropy loss Together they serve as the total loss function during model training, expressed as: ； Where, is the hyperparameter, that is, the weight of the margin-based negative sample loss.

10. A personalized EEG emotion recognition system based on dynamic brain region division, characterized by: include: The data preprocessing module is configured to: preprocess the EEG signal and construct an EEG signal matrix; The emotion intensity attention module is configured to: extract the five-band differential entropy eigenvalues ​​based on the preprocessed EEG signal, perform multi-head self-attention calculations in combination with rotational position encoding, and output channel feature representations with strong emotion intensity; The functional space partitioning module is configured to: dynamically divide brain functional areas using the CNM algorithm based on the constructed EEG signal matrix, optimize the brain functional area structure with modularity value, and output brain area features; The emotion prediction module is configured to: fuse channel feature representation and brain region features, predict emotion labels through graph neural networks, and output emotion recognition results.

Images