Methods and systems for extracting EEG signal features using brain region channel exchange data enhancement

Abstract

This invention discloses a method and system for extracting EEG signal features using brain region channel exchange data enhancement. The invention includes: S1, acquiring EEG signals from multiple subjects under different stimuli; S2, selecting EEG signals from different subjects under the same stimulus, grouping and pairing them, and then performing cross-data enhancement on some brain region channels to construct positive and negative sample pairs; S3, using a group projector and contrast loss function to backpropagate and train a basic encoder based on the positive and negative sample pairs; S4, constructing an EEG signal classification model using the backpropagation-trained basic encoder and classifier, and fine-tuning the EEG signal classification model by constructing EEG signals and classification labels from finite label sample pairs, thereby obtaining a basic encoder that can be used for EEG signal feature extraction. This invention aims to extract deep spatiotemporal features of EEG signals, improve the robustness and generalization of features, and enhance the performance of cross-subject classification tasks.

Description

Technical Field

[0001] This invention belongs to the field of electroencephalogram (EEG) signal processing technology, specifically relating to a method and system for extracting EEG signal features by enhancing brain region channel exchange data. Background Technology

[0002] With the continuous upgrading of acquisition technology and the ongoing optimization of analysis methods, electroencephalography (EEG) signals have become an extremely important research tool in the field of brain science. In the area of ​​cognitive state assessment, EEG signals have shown extremely broad and valuable application prospects. However, EEG signals face many significant challenges in practical applications. On the one hand, EEG signals contain numerous random interference factors, such as whether the subject is distracted during task performance or their own fatigue state; these factors increase signal noise to varying degrees. On the other hand, there are unavoidable differences in EEG signals between individuals, and these inter-individual differences bring significant difficulties and challenges to maximizing cross-subject similarity in contrastive learning. Meanwhile, deep learning requires a large amount of labeled data to support its learning and optimization process. However, obtaining large-scale labeled EEG signals faces many practical difficulties and obstacles. Labeling EEG signals is not a simple operation; it requires the involvement of experts with solid neurophysiological knowledge and a deep understanding of EEG data. The high cost of labeling and the high professional knowledge threshold make constructing high-quality, large-scale EEG datasets extremely challenging. Furthermore, different laboratories often employ different experimental designs, stimulus materials, and data acquisition procedures when conducting research. This difference results in significant variations in the format and characteristics of data generated from different studies, hindering effective joint training and further impeding cross-laboratory data integration and collaborative research. EEG data itself also suffers from spatiotemporal redundancy, meaning that data from adjacent time points and spatial channels may exhibit high correlations. This redundancy can lead to overfitting during model training and significantly reduce the model's generalization ability, thus affecting its performance and reliability in practical applications. Overcoming these challenges effectively and further improving the performance of EEG-based feature extraction and classification has become a key research objective and a crucial problem urgently needing to be solved in this field. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a method and system for extracting EEG signal features by enhancing brain region channel exchange data, in order to address the above-mentioned problems of the prior art. The present invention aims to extract deep spatiotemporal features of EEG signals, improve the robustness and generalization of features, and enhance the performance of cross-subject classification tasks.

[0004] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0005] A method for extracting EEG signal features enhanced by brain region channel exchange data, comprising the following steps:

[0006] S1, acquire EEG signals from multiple subjects under different stimuli;

[0007] S2, select different subjects' EEG signals under the same stimulus, group and pair them, and then perform cross-data enhancement on some brain regions to construct positive and negative sample pairs;

[0008] S3, the basic encoder is backpropagated and trained using a group projector and a contrastive loss function based on positive and negative sample pairs;

[0009] S4 constructs an EEG signal classification model using the basic encoder and classifier after backpropagation training. The EEG signal classification model is fine-tuned by constructing EEG signals and classification labels in a limited number of labeled sample pairs, thereby obtaining a basic encoder that can be used for EEG signal feature extraction.

[0010] Optionally, step S2 includes:

[0011] S2.1, for all Each stimulus, for each stimulus All of the following The EEG signals of each subject were aggregated into one EEG signal sample group. Finally obtained A group of EEG signal samples ;

[0012] S2.2, for each EEG signal sample group , will the internal Random brain region channel crossover was performed on the EEG signals of each subject, and the EEG signals before and after the random brain region channel crossover were used to form positive sample pairs; targeting A set of original EEG signal samples and The EEG signal sample group consists of random brain region channel crossover transformations. A set of EEG signal samples, any set of EEG signal samples and the EEG signal sample groups obtained before and after random brain region channel crossover. The subject's EEG signal and two other (P-1) EEG signal sample groups constituted a negative sample pair.

[0013] Optionally, in step S2.2, the internal Random brain region channel crossover of the EEG signals of each subject includes:

[0014] S2.2.1, Perform individual allocation: internally EEG signals were randomly selected from the EEG signals of each subject for pairing. ,in and respectively stimuli The following EEG signal sample group The first in and The brain signals of the subjects, among which The value range is 1~ Random numbers between;

[0015] S2.2.2, Perform channel switching: Pair the selected EEG signals. Randomly select a brain region channel for EEG signal cross-transformation, so that the EEG signal... Selected brain region pathways and EEG signals A new brain signal is formed in the brain region other than the selected brain region. Selected brain region pathways and EEG signals A new brain signal is formed in the brain region channel other than the selected brain region channel;

[0016] S2.2.3, Perform separation and recombination: Randomly divide all the obtained new EEG signals into two groups, with the same EEG signal paired into different groups, ultimately obtaining two EEG signal sample groups sharing the same stimulus. and .

[0017] Optionally, the basic encoder consists of a convolutional layer, a max pooling layer, a multi-level encoding module, and an average pooling layer connected in sequence. The encoding module is composed of multi-level convolutional layers, and the input features of the encoding module and the final output features of the multi-level convolutional layers are skip-connected to serve as the output features of the encoding module. The convolutional layer consists of convolution, batch normalization, and ReLU linear activation unit functions.

[0018] Optionally, in step S3, when backpropagating the basic encoder using a group projector and a preset contrastive loss function, the process includes inputting the latent representations of individual subjects obtained from the basic encoder into the group projector to obtain group features in the latent space, and using the contrastive loss function shown in the following equation. To train the basic encoder using backpropagation:

[0019] ;

[0020] in, for and Similarity between them for and Similarity between them for and Similarity between them EEG signal sample group Samples in Potential representations, EEG signal sample group Samples in Potential representations, EEG signal sample group Samples in Potential representations, EEG signal sample group Samples in Potential representations, For the sample size, It is a temperature constant. The selection function for determining positive and negative sample pairs, when It is 1 if it is true, otherwise it is 0. and This represents two sets of EEG signal samples that share a common stimulus.

[0021] Optionally, in step S3, the group projector includes a first multilayer perceptron (MLP) and a one-dimensional max-pooling layer. The first MLP has a three-layer structure, and each of the three layers consists of a fully connected module, a batch normalization module, and a ReLU linear activation unit. The number of neurons in the three fully connected modules of the first MLP are 1024, 2048, and 4096, respectively. The one-dimensional max-pooling layer is used to process the output of the MLP along 4096 feature dimensions. The maximum value of each upgraded feature is taken to obtain the population features in the latent space.

[0022] Optionally, in step S4, when constructing an EEG signal classification model from the basic encoder and classifier after backpropagation training, the classifier used includes a second multilayer perceptron (MLP) and a Softmax activation function module. The second MLP is a three-layer structure, and each of the three layers consists of a fully connected module, a batch normalization module, and a ReLU linear activation unit. The number of neurons in the three fully connected modules of the second MLP are 512, 256, and 128, respectively, and the last layer has a random loss module to implement random loss of features. The Softmax activation function module is used to classify the second MLP to obtain the classification label of the EEG signal.

[0023] Furthermore, this embodiment also provides a brain region channel exchange data-enhanced EEG signal feature extraction system, including a microprocessor and a memory interconnected, wherein the microprocessor is programmed or configured to execute the brain region channel exchange data-enhanced EEG signal feature extraction method.

[0024] Furthermore, this embodiment also provides a computer-readable storage medium storing a computer program or instructions that are programmed or configured to execute, via a processor, the brain region channel exchange data-enhanced EEG signal feature extraction method.

[0025] In addition, this embodiment also provides a computer program product, including a computer program or instructions, which are programmed or configured to execute, via a processor, the brain region channel exchange data-enhanced EEG signal feature extraction method.

[0026] Compared with existing technologies, the present invention mainly achieves the following beneficial effects: The brain region channel exchange data enhancement method for EEG signal feature extraction is a self-supervised learning method. It employs a contrastive learning framework, performing data enhancement by cross-exchanging corresponding brain region channels after pairing, while maintaining stimulus features unchanged. This method allows contrastive learning to better utilize the stimulus alignment characteristics within EEG signal groups, ultimately enabling the model to learn key representations and achieve significant performance improvements in downstream tasks. When receiving the same stimulus, EEG signals in some identical brain regions are very similar; therefore, the method of the present invention can effectively extract deep spatiotemporal features of EEG signals. The present invention selects signals from corresponding brain region channels of different subjects at the same time segment for cross-data enhancement, forming more positive and negative sample pairs, and extracts key spatiotemporal features of EEG signals through contrastive learning. The present invention can be used for feature extraction of EEG signal data from various EEG paradigms, and is more effective for datasets with limited labels, providing support for various tasks based on EEG signal feature extraction. This invention fully utilizes the similarity of EEG signals in the same brain region for the same stimulus. By exchanging signals from the same brain region channels among different subjects for data augmentation, contrastive learning can extract spatiotemporal features of EEG, improving the robustness and generalization of these features. Furthermore, by extracting features from the same group including different subjects, the brain region-specific features extracted from multiple subjects are more generalizable, thus improving performance on cross-subject classification tasks. This invention's method achieves data augmentation and enhances performance on cross-subject classification tasks. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the basic process of the method in an embodiment of the present invention.

[0028] Figure 2 This is a schematic diagram illustrating the basic principle of the method in an embodiment of the present invention.

[0029] Figure 3 This is a schematic diagram of subject sampling selection in an embodiment of the present invention.

[0030] Figure 4 This is a schematic diagram of a brain region channel exchange data enhancement method according to an embodiment of the present invention.

[0031] Figure 5 This is a schematic diagram of the network structure of the basic encoder in an embodiment of the present invention.

[0032] Figure 6 This is a schematic diagram of the network structure of the group projector in an embodiment of the present invention.

[0033] Figure 7 This is a schematic diagram of the network structure of the classifier in an embodiment of the present invention.

[0034] Figure 8 This is a confusion matrix used in this embodiment of the invention to evaluate the mixed subject emotion classification on a four-class classification problem.

[0035] Figure 9 This is a confusion matrix used in embodiments of the present invention to evaluate the confusion of mixed subject emotion classification on the classification results. Detailed Implementation

[0036] This invention is based on the fact that EEG signals in some identical brain regions are very similar when receiving the same stimulus, thus allowing the extraction of deep spatiotemporal features of EEG signals. By selecting signals from corresponding brain regions of different subjects at the same time segment and performing cross-data enhancement, more positive and negative sample pairs are formed. Through comparative learning, key spatiotemporal features of EEG signals are extracted. To enable those skilled in the art to better understand the technical solution of this invention, the following will provide a more detailed description of the technical solution in conjunction with the accompanying drawings of the embodiments of this invention.

[0037] like Figure 1 and Figure 2 As shown, the method for extracting EEG signal features using brain region channel exchange data enhancement in this embodiment includes the following steps:

[0038] S1, acquire EEG signals from multiple subjects under different stimuli;

[0039] S2, select different subjects' EEG signals under the same stimulus, group and pair them, and then perform cross-data enhancement on some brain regions to construct positive and negative sample pairs;

[0040] S3, the basic encoder is backpropagated and trained using a group projector and a contrastive loss function based on positive and negative sample pairs;

[0041] S4 constructs an EEG signal classification model using the basic encoder and classifier after backpropagation training. The EEG signal classification model is fine-tuned by constructing EEG signals and classification labels in a limited number of labeled sample pairs, thereby obtaining a basic encoder that can be used for EEG signal feature extraction.

[0042] In step S1 of this embodiment, the acquisition of EEG signals from multiple subjects under different stimuli can be performed using the required EEG signal preprocessing method. For example, as an optional implementation, the preprocessing steps include: denoising the EEG signals; dividing the denoised EEG signals into windows of fixed-size time; and dividing the electrodes into regions according to the 10-20 system based on brain functional partitions and task-related prior knowledge. The denoising can be performed using the required denoising method. For example, as an optional implementation, the denoising steps include: downsampling the EEG signals; baseline correction of the downsampled EEG signals by subtracting the channel mean; and removing noise interference from the baseline-corrected EEG signals using a bandpass filter. When dividing the electrodes into regions according to the 10-20 system based on brain functional partitions and task-related prior knowledge, the initial division can be based on the main brain regions—frontal lobe, parietal lobe, occipital lobe, and temporal lobe.

[0043] In this embodiment, step S2 includes:

[0044] S2.1, as Figure 3 As shown, randomly select from the current training rounds. One unsampled video clip: As a stimulus; then select 100 subjects. For all... Each stimulus, for each stimulus All of the following The EEG signals of each subject were aggregated into one EEG signal sample group. Finally obtained A group of EEG signal samples ;

[0045] S2.2, for each EEG signal sample group , will the internal The EEG signals of each subject were subjected to random brain region channel crossover, and the EEG signals before and after the random brain region channel crossover were used to form positive sample pairs; such as Figure 4 As shown, for the grouped data, each pair of participants within a group underwent random exchange of corresponding brain region channels. The resulting EEG signals exhibited different signal characteristics from each participant, forming positive sample pairs. A set of original EEG signal samples and The EEG signal sample group consists of random brain region channel crossover (Exchange) A set of EEG signal samples, any set of EEG signal samples and the EEG signal sample groups obtained before and after random brain region channel crossover. The subject's EEG signal and two other (P-1) EEG signal sample groups constituted a negative sample pair.

[0046] In this embodiment, step S2.2 will internally... Random brain region channel crossover of the EEG signals of each subject includes:

[0047] S2.2.1, Perform individual allocation: internally EEG signals were randomly selected from the EEG signals of each subject for pairing. ,in and respectively stimuli The following EEG signal sample group The first in and The brain signals of the subjects, among which The value range is 1~ Random numbers between;

[0048] S2.2.2, Perform channel switching: Pair the selected EEG signals. Randomly select a brain region channel for EEG signal cross-transformation, so that the EEG signal... Selected brain region pathways and EEG signals A new brain signal is formed in the brain region other than the selected brain region. Selected brain region pathways and EEG signals A new brain signal is formed in the brain region channel other than the selected brain region channel;

[0049] S2.2.3, Perform separation and recombination: Randomly divide all the obtained new EEG signals into two groups, with the same EEG signal paired into different groups, ultimately obtaining two EEG signal sample groups sharing the same stimulus. and , can be represented as:

[0050] ;

[0051] in, This represents random brain region pathway crossover.

[0052] The basic encoder in this embodiment is an improved version of the one-dimensional residual neural network (ResNet18-1D) architecture. For example... Figure 5As shown, in this embodiment, the basic encoder consists of sequentially connected convolutional layers, max-pooling layers, multi-level encoding modules (specifically 8 levels in this embodiment), and average pooling layers. The encoding modules are composed of multi-level convolutional layers, and the input features of the encoding modules and the final output features of the multi-level convolutional layers are skip-connected to serve as the output features of the encoding modules. The convolutional layers consist of convolution, batch normalization, and ReLU linear activation units. By introducing a residual learning mechanism, the degradation problem in deep neural network training is solved. The one-dimensional convolutional layers can efficiently capture local spatiotemporal features in EEG signals, and the design of residual blocks also helps the network learn more complex feature representations while maintaining computational efficiency.

[0053] In step S3 of this embodiment, when backpropagating the basic encoder using a group projector and a preset contrastive loss function, the process includes inputting the latent representation of a single subject obtained from the basic encoder into the group projector to obtain group features in the latent space; and inputting the EEG feature representation of the entire group of subjects into the group projector to obtain group features. The group features are then input into the loss function to calculate the loss. The cosine similarity between the two representations is used as the probability output, and the loss function uses a temperature-scaled cross-entropy function. Specifically, in this embodiment, the contrastive loss function shown in the following formula is used. To train the basic encoder using backpropagation:

[0054] ;

[0055] in, for and Similarity between them for and Similarity between them for and Similarity between them EEG signal sample group Samples in Potential representations, EEG signal sample group Samples in Potential representations, EEG signal sample group Samples in Potential representations, EEG signal sample group Samples in Potential representations, For the sample size, It is a temperature constant. The selection function for determining positive and negative sample pairs, when It is 1 if it is true, otherwise it is 0. and This represents two sets of EEG signal samples that share a common stimulus.

[0056] like Figure 6 As shown, in step S3 of this embodiment, the group projector includes a first multilayer perceptron (MLP) and a one-dimensional max-pooling layer. The first MLP has a three-layer structure, and each of the three layers consists of a fully connected module, a batch normalization module, and a ReLU linear activation unit. The number of neurons in the three fully connected modules of the first MLP are 1024, 2048, and 4096, respectively. The one-dimensional max-pooling layer is used to process the output of the MLP along the 4096 feature dimensions. The maximum value of each upgraded feature is taken to obtain the population features in the latent space.

[0057] like Figure 7 As shown, in step S4 of this embodiment, when constructing the EEG signal classification model using the basic encoder and classifier after backpropagation training, the classifier used includes a second multilayer perceptron (MLP) and a Softmax activation function module. The second MLP has a three-layer structure, and each of the three layers consists of a fully connected module, a batch normalization module, and a ReLU linear activation unit. The number of neurons in the three fully connected modules of the second MLP are 512, 256, and 128, respectively, and the last layer has a random loss module to implement random loss of features (in this embodiment, the probability is 0.5, which helps the model learn more robust features). The Softmax activation function module is used to classify the second MLP to obtain the classification label of the EEG signal, such as emotion labels (negative, neutral, and positive).

[0058] To validate the brain region channel exchange data-enhanced EEG signal feature extraction method of this embodiment, the publicly available DEAP and SEED datasets were used for in-subject and mixed-subject trials, respectively. The DEAP dataset includes 32-channel EEG signals and 8-channel peripheral physiological signals recorded by 32 subjects while watching 40 one-minute music videos. Each experimental data segment included a 3-second resting state and a 60-second stimulation state. The SEED dataset records in detail the EEG signals of 15 subjects when receiving three types of emotional stimuli (positive, neutral, and negative). The EEG signals were recorded through 62 electrodes at a high sampling rate of 1000Hz, and the stimuli came from 15 different video clips they watched. In this embodiment, on the DEAP dataset, the method of this embodiment was first compared with three supervised methods on two dimensions of emotion (i.e., valence and arousal): a dense convolutional network (CDCN) with channel fusion, a residual long short-term memory network (MMResLSTM) utilizing multimodal data, and a hybrid network of convolutional neural networks and recurrent neural networks with channel attention mechanism (ACRNN). To verify the effectiveness of the proposed method in data augmentation and self-supervised learning, it was further compared with the GAN-based data augmentation method (MCLFS-GAN) and the GAN-based self-supervised data augmentation framework (GANSER). The proposed method employed two training approaches for comparison: fully-supervised learning and fine-tuned training. Fully-supervised learning involved using the complete model without pre-training, while fine-tuned training involved pre-training the complete model followed by fine-tuning optimization. The final results are shown in Table 1.

[0059] Table 1: Classification results of mixed subjects on the DEAP dataset

[0060]

[0061] As shown in Table 1, compared with existing supervised methods and data augmentation methods, the models obtained by the supervised learning and fine-tuning training methods in this embodiment have superior performance in terms of valence, arousal, and four-class classification. In particular, the accuracy of the model obtained after fine-tuning is 3.49%, 3.32%, and 4.97% higher than the fully supervised baseline, respectively. This indicates that the self-supervised process of this embodiment has a significant improvement effect on downstream tasks.

[0062] In this embodiment, on the SEED dataset, our method is first compared with current state-of-the-art fully supervised methods, including: GRSLR (using a graph-regularized sparse linear regression model); BiHDM (using two independent recurrent networks to process EEG signals from the two hemispheres of the brain); DGCNN (using a dynamic graph convolutional neural network); and ResNet18 (a residual network based on a one-dimensional convolutional neural network). The final results are shown in Table 2.

[0063] Table 2: Mixed subject emotion classification results on the SEED dataset

[0064]

[0065] As shown in Table 2, the method in this embodiment outperforms four advanced model networks: GRSLR, DGCNN, BiHDM, and ResNet18 1D kernel. The supervised learning model obtained by this embodiment achieves an average accuracy of 89.73%, and the fine-tuned model achieves an average accuracy of 93.94%. When fine-tuning using labeled samples from the complete training set, the fine-tuned model exceeds the average accuracy of the fully supervised learning model by 4.27%. This also demonstrates that the model in this embodiment significantly improves downstream tasks by utilizing a large amount of unlabeled data.

[0066] To investigate the impact of the number of samples per group (Q) and the number of video segments selected in each iteration (P) on contrastive learning, the performance comparison results of different hyperparameters on the DEAP and SEED datasets are shown in Table 3 in this embodiment.

[0067] Table 3: Performance comparison of different hyperparameters on two datasets

[0068]

[0069] In Table 3, Q represents the number of samples per group, and P represents the number of video segments. Table 3 shows that different combinations of hyperparameters significantly affect the performance of the contrastive learning algorithm, with opposite trends observed in the DEAP and SEED datasets. Given a larger Q, a smaller P is suitable for DEAP, while a larger P is preferred for SEED. This may be due to the different labeling methods in the two datasets. In the SEED dataset, emotion labels are assigned by the experiment designers based on the emotional attributes of the video stimuli. In the DEAP dataset, emotion labels are determined by the participants' ratings, a labeling method more related to individual differences among participants than to the stimuli themselves, as in SEED. Furthermore, a larger P increases the difficulty of contrastive learning. In this case, the model tends to focus on extracting stimulus-related features, ignoring personalized features unrelated to the stimulus. Therefore, a larger P leads to better results in SEED, but hinders better results in DEAP. Additionally, it can be observed that, generally, with P remaining constant, a larger Q leads to higher pre-training accuracy. This may be because larger groups contain more comprehensive group-level stimulus-related features, helping to mitigate the effects of random interference, fatigue, and individual differences. However, high accuracy during pre-training is not always beneficial for emotion recognition. An excessively large Q-value can cause the model to focus too much on the aggregation of group-level stimulus-related features, leading the base encoder to neglect learning some emotion-related features and thus hindering better emotion recognition. Therefore, choosing an appropriate Q-value is crucial for constructing group-based contrastive learning. Thus, in the implementation of this method, selecting an appropriate value based on the dataset is essential. This choice should aim to capture diverse and comprehensive group-level features while also focusing on the subtle individual differences necessary for accurate emotion recognition. The optimal group size should allow the model to learn representations that effectively generalize to all participants while preserving necessary task-relevant features.

[0070] In this embodiment, an ablation experiment was conducted. Based on the brain region channel exchange data enhancement EEG signal feature extraction method of this embodiment, four models were designed: non-group, non-augment, fully-supervised, and fine-tuned. Non-group was achieved by removing group samples, non-augment was achieved by removing brain region exchange data enhancement, fully-supervised was achieved by using the complete model but without pre-training and directly performing supervised training, and fine-tuned was achieved by using the complete model for pre-training and subsequent fine-tuning optimization. The final results are shown in Table 4.

[0071] Table 4: Ablation Experiment Results

[0072]

[0073] As shown in Table 4, the non-grouping method significantly reduced emotion recognition performance by more than 1.5% on both the DEAP and SEED datasets. The non-augment method resulted in a significant decrease in accuracy of over 3% on the DEAP dataset and over 1.2% on the SEED dataset. This verifies that data augmentation methods can provide more unlabeled data to aid in the pre-training of self-supervised learning tasks, playing a crucial role in the algorithm framework and significantly improving the performance of downstream emotion classification tasks. The fully-supervised model was compared with a fully supervised model of the same structure, achieving 89.73% accuracy on the SEED dataset and 87.68% accuracy on the DEAP dataset. The fine-tuned model achieved 93.94% accuracy on the SEED dataset and 92.42% accuracy on the DEAP dataset, showing a significant improvement in accuracy compared to the fully supervised baseline. This indicates that the self-supervised process of the method in this embodiment has a significant improvement effect on downstream emotion recognition tasks.

[0074] In this embodiment, a confusion matrix evaluation of mixed subject emotion classification was performed. Figure 8 This is the confusion matrix used in this embodiment to evaluate the mixed subject emotion classification on the four-class classification problem. Figure 9 This example uses a confusion matrix to evaluate the impact of mixed subject emotion classification on the classification results. See also... Figure 8 As shown in the confusion matrix on the DEAP dataset for the four-class classification problem, the method in this embodiment achieves good performance in each class of the four-class classification problem (high arousal and high valence, high arousal and low valence, low arousal and high valence, and low arousal and low valence), especially in the low arousal and high valence classes. See also Figure 9 As shown in the confusion matrix of the classification results, the method of this embodiment achieved good accuracy in all three categories (negative, neutral, and positive), especially outperforming negative and neutral emotions in positive emotions, which is consistent with the conclusions of existing research.

[0075] As can be seen, the EEG signal feature extraction method enhanced by brain region channel exchange data in this embodiment addresses the problems of deep spatiotemporal feature redundancy, strong inter-individual specificity, and limited dataset labels in EEG signals. It utilizes the idea of ​​contrastive learning and incorporates the construction of positive and negative samples through random brain region channel exchange of EEG signals under the same stimulus to improve the self-supervised training effect. This EEG signal feature extraction method enhanced by brain region channel exchange data can be used in specific tasks of emotion recognition. It achieved superior performance in mixed-subject emotion recognition experiments, exhibiting advantages such as high evaluation accuracy, good robustness, and strong generalization performance. This method can be used for feature extraction of various EEG paradigms, and is more effective for datasets with limited labels, providing support for various tasks based on EEG signal feature extraction.

[0076] Furthermore, this embodiment also provides a brain region channel exchange data-enhanced EEG signal feature extraction system, including a microprocessor and a memory interconnected, wherein the microprocessor is programmed or configured to execute the brain region channel exchange data-enhanced EEG signal feature extraction method.

[0077] Furthermore, this embodiment also provides a computer-readable storage medium storing a computer program or instructions that are programmed or configured to execute, via a processor, the brain region channel exchange data-enhanced EEG signal feature extraction method.

[0078] In addition, this embodiment also provides a computer program product, including a computer program or instructions, which are programmed or configured to execute, via a processor, the brain region channel exchange data-enhanced EEG signal feature extraction method.

[0079] 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.

[0080] 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 extracting features from electroencephalogram (EEG) signals enhanced by brain region channel exchange data, characterized in that, Includes the following steps: S1, acquire EEG signals from multiple subjects under different stimuli; S2, select different subjects' EEG signals under the same stimulus, group and pair them, and then perform cross-data enhancement on some brain regions to construct positive and negative sample pairs; S3, the basic encoder is backpropagated and trained using a group projector and a contrastive loss function based on positive and negative sample pairs; S4. After the basic encoder and classifier have been trained by backpropagation, a brain signal classification model is constructed. The brain signal classification model is fine-tuned by constructing brain signals and classification labels in a limited number of labeled sample pairs, thereby obtaining a basic encoder that can be used for brain signal feature extraction. Step S2 includes: S2.1, for all Each stimulus, for each stimulus All of the following The EEG signals of each subject were aggregated into one EEG signal sample group. Finally obtained A group of EEG signal samples ; S2.2, for each EEG signal sample group , will the internal Random brain region channel crossover was performed on the EEG signals of each subject, and the EEG signals before and after the random brain region channel crossover were used to form positive sample pairs; targeting A set of original EEG signal samples and The EEG signal sample group consists of random brain region channel crossover transformations. A set of EEG signal samples, any set of EEG signal samples and the EEG signal sample groups obtained before and after random brain region channel crossover. The EEG signals of the subjects in the sample and the other 2 (P-1) EEG signal sample groups constitute a negative sample pair; In step S2.2, the internal Random brain region channel crossover of the EEG signals of each subject includes: S2.2.1, Perform individual allocation: internally EEG signals were randomly selected from the EEG signals of each subject for pairing. ,in and respectively stimuli The following EEG signal sample group The first in and The brain signals of the subjects, among which The value range is 1~ Random numbers between; S2.2.2, Perform channel switching: Pair the selected EEG signals. Randomly select a brain region channel for EEG signal cross-transformation, so that the EEG signal... Selected brain region pathways and EEG signals A new brain signal is formed in the brain region other than the selected brain region. Selected brain region pathways and EEG signals A new brain signal is formed in the brain region channel other than the selected brain region channel; S2.2.3, Perform separation and recombination: Randomly divide all the obtained new EEG signals into two groups, with the same EEG signal paired into different groups, ultimately obtaining two EEG signal sample groups sharing the same stimulus. and .

2. The method for extracting EEG signal features by enhancing brain region channel exchange data according to claim 1, characterized in that, The basic encoder consists of a convolutional layer, a max pooling layer, a multi-level encoding module, and an average pooling layer connected in sequence. The encoding module is composed of multi-level convolutional layers, and the input features of the encoding module and the final output features of the multi-level convolutional layers are skip-connected to serve as the output features of the encoding module. The convolutional layer consists of convolution, batch normalization, and ReLU linear activation unit functions.

3. The method for extracting EEG signal features by enhancing brain region channel exchange data according to claim 1, characterized in that, In step S3, when backpropagating the basic encoder using a group projector and a preset contrastive loss function, the process includes inputting the latent representation of a single subject obtained from the basic encoder into the group projector to obtain group features in the latent space, and using the contrastive loss function shown in the following equation. To train the basic encoder using backpropagation: ； in, for and Similarity between them for and Similarity between them for and Similarity between them EEG signal sample group Samples in Potential representations, EEG signal sample group Samples in Potential representations, EEG signal sample group Samples in Potential representations, EEG signal sample group Samples in Potential representations, For the sample size, It is a temperature constant. The selection function for determining positive and negative sample pairs, when It is 1 if it is true, otherwise it is 0. and This represents two sets of EEG signal samples that share a common stimulus.

4. The method for extracting EEG signal features by enhancing brain region channel exchange data according to claim 1, characterized in that, In step S3, the group projector includes a first multilayer perceptron (MLP) and a one-dimensional max-pooling layer. The first MLP has a three-layer structure, and each of the three layers consists of a fully connected module, a batch normalization module, and a ReLU linear activation unit. The number of neurons in the three fully connected modules of the first MLP are 1024, 2048, and 4096, respectively. The one-dimensional max-pooling layer is used to process the output of the MLP along 4096 feature dimensions. The maximum value of each upgraded feature is taken to obtain the population features in the latent space.

5. The method for extracting EEG signal features by enhancing brain region channel exchange data according to claim 1, characterized in that, In step S4, when constructing the EEG signal classification model using the basic encoder and classifier after backpropagation training, the classifier used includes a second multilayer perceptron (MLP) and a Softmax activation function module. The second MLP has a three-layer structure, and each of the three layers consists of a fully connected module, a batch normalization module, and a ReLU linear activation unit. The number of neurons in the three fully connected modules of the second MLP are 512, 256, and 128, respectively, and the last layer has a random loss module to implement random loss of features. The Softmax activation function module is used to classify the second MLP to obtain the classification label of the EEG signal.

6. A brain region channel exchange data enhancement system for extracting EEG signal features, comprising a microprocessor and a memory interconnected thereon, characterized in that, The microprocessor is programmed or configured to perform the brain region channel exchange data enhancement method for extracting EEG signal features according to any one of claims 1 to 5.

7. 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, via a processor, the brain region channel exchange data enhancement method for extracting EEG signal features according to any one of claims 1 to 5.

8. A computer program product, comprising a computer program or instructions, characterized in that, The computer program or instructions are programmed or configured to execute, via a processor, the brain region channel exchange data enhancement method for extracting EEG signal features according to any one of claims 1 to 5.

Images