Method and system for few-channel electroencephalogram emotion recognition based on physiological prior guidance and cross-channel knowledge transfer

Abstract

The application discloses a few-channel electroencephalogram emotion recognition method and system based on physiological priori and cross-channel knowledge transfer, and belongs to the field of electroencephalogram signal processing. The application aims at the problems of spatial topology loss, cross-channel distribution deviation and full-channel model deployment, constructs full-channel and few-channel paired samples, combines a physiological priori graph convolution module, an emotion anchor point aggregation module and a cross-channel knowledge transfer mechanism, and realizes high-reliability emotion recognition under few channels. The physiological priori graph convolution fuses electrode distance, hemisphere symmetry and dynamic connection, compensates for the loss of spatial information; the emotion anchor point aggregation enhances dynamic representation through learnable anchor points, subspace attention, gating and residual weighting; the cross-channel transfer uses the global features, class prototypes and soft label constraints of the teacher network to make the student network approach the full-channel performance. The application reduces the cost and complexity of the device, improves the recognition accuracy, robustness and generalization ability, and is suitable for portable electroencephalogram emotion recognition scenes.

Description

Technical Field

[0001] This invention relates to the field of electroencephalogram (EEG) signal processing, and in particular to a method and system for few-channel EEG emotion recognition based on physiological prior guidance and cross-channel knowledge transfer. Background Technology

[0002] Electroencephalography (EEG) can non-invasively reflect neural activity in the cerebral cortex and has been widely used in applications such as emotion recognition, human-computer interaction, mental health monitoring, and smart cockpits. Current high-precision EEG emotion recognition methods typically rely on high-density, full-channel EEG devices with 62 or 32 channels. While these devices can acquire relatively complete spatial topological information, they suffer from drawbacks such as large device size, complex wearing process, high cost, poor user compliance, and difficulty in adapting to everyday dynamic scenarios.

[0003] Due to their advantages such as lightweight design, low power consumption, and ease of deployment, few-channel EEG devices are gradually becoming an important hardware form for wearable emotion-based brain-computer interfaces. However, few-channel EEG devices still face the following technical challenges in emotion recognition tasks: First, the reduced number of electrodes leads to sparse spatial information, making it difficult to fully model the whole-brain collaborative relationships between the prefrontal, parietal, temporal, and occipital lobes; second, there are significant distributional offsets among different channel configurations, device forms, and subjects, making it difficult to directly transfer laboratory full-channel models to few-channel scenarios; third, few-channel scenarios lack redundant channels for spatial filtering, making signals more susceptible to electrode displacement, electromyography artifacts, and environmental noise.

[0004] Existing methods for few-channel EEG emotion recognition mainly include fixed-channel selection methods based on neurophysiological priors, channel screening methods based on statistical indicators or deep models, and cross-device transfer learning methods. Fixed-channel selection methods offer some interpretability but struggle to adapt to individual differences; purely data-driven methods are prone to black-boxing and overfitting; while cross-device transfer learning methods can mitigate inter-domain differences, they often lack systematic compensation for spatial topological gaps at different channel resolutions. Therefore, a solution for EEG emotion recognition that simultaneously considers few-channel deployability, cross-channel knowledge transfer capability, and physiological interpretability is needed. Summary of the Invention

[0005] This invention provides a method and system for few-channel EEG emotion recognition based on physiological prior guidance and cross-channel knowledge transfer, aiming to solve the problems of missing spatial topological information, insufficient generalization ability of few-channel models, difficulty in directly deploying full-channel algorithms to portable EEG devices, and obvious cross-channel distribution offset in existing few-channel EEG emotion recognition.

[0006] On the one hand, this invention provides a few-channel EEG emotion recognition method based on physiological prior guidance and cross-channel knowledge transfer, which includes the following steps:

[0007] Step 1: Construct paired EEG feature samples with full-channel and few-channel configurations;

[0008] Based on the input full-channel EEG features, corresponding electrode channels are extracted according to the target few-channel configuration to obtain few-channel EEG features; the full-channel EEG features and few-channel EEG features are kept in a one-to-one correspondence in terms of subject identity, experimental session, stimulus segment, time window and emotional label to form paired EEG feature samples.

[0009] Step 2: Extract spatial-spectral feature sequences based on the physiological prior map convolution module;

[0010] A physiological prior adjacency matrix was constructed based on the distance relationship between EEG electrodes in the scalp space and the symmetrical relationship between electrodes in the left and right hemispheres. Set up a learnable dynamic adjacency matrix. To capture data-driven channel connectivity; and to matrix and The fusion yields a hybrid adjacency matrix. ;

[0011] Based on hybrid adjacency matrix Normalized and scaled Laplacian matrices are constructed, and spatial-spectral feature sequences of EEG signals are extracted by Chebyshev plot convolution.

[0012] Step 3: Extract global sentiment representation based on the sentiment anchor aggregation module;

[0013] An emotion anchor aggregation module is set up in both the full-channel teacher network and the few-channel student network; the full-channel teacher network receives full-channel EEG features, and extracts the full-channel spatial-spectral feature sequence through the physiological prior map convolution module. The student network receives a small number of EEG features, which are then extracted using a physiological prior graph convolution module to obtain a small number of spatial-spectral feature sequences. ;

[0014] The spatial-spectral feature sequence is input into the sentiment anchor aggregation module. Through learnable sentiment anchors, subspace attention matching, gating adjustment, and residual weighted aggregation, a global sentiment representation is obtained. The global sentiment representation is then input into a classifier to output the sentiment category probability.

[0015] Step 4: Conduct cross-channel knowledge transfer training based on the full-channel teacher network and the few-channel student network;

[0016] Full-channel global sentiment representation obtained using teacher network and sentiment anchor aggregation module The system uses category prototypes or soft labels to supervise and constrain student networks, and combines teacher classification loss, student classification loss, prototype alignment loss and knowledge distillation loss to achieve cross-channel knowledge transfer training. This enables student networks with fewer channels to acquire sentiment discrimination capabilities close to those of full-channel teacher networks with a reduced number of electrodes.

[0017] Step 5: Perform online inference based on the trained few-channel student network;

[0018] During the online inference phase, the few-channel EEG features to be identified are input into the trained few-channel student network. After processing by the emotion anchor aggregation module and classifier, the probability of the emotion category is output, and the final emotion recognition result is determined based on the category with the highest probability.

[0019] Furthermore, in step 1, paired EEG feature samples with full-channel and few-channel characteristics are constructed using publicly available EEG emotion datasets or self-collected EEG emotion datasets.

[0020] Furthermore, in step 1, the tensor dimension of the full-channel EEG features / few-channel EEG features is... ,in, Indicates the number of time windows. The number of channels representing brainwave signals. This represents the dimension of the frequency domain features extracted for each channel.

[0021] Furthermore, in step 2, the physiological prior adjacency matrix for:

[0022]

[0023]

[0024] in, For the first The first electrode and the second The distance between the electrodes For the first The first electrode and the second The hemispherical symmetry of each electrode indicates the quantity. For distance scale parameters, The symmetry enhancement coefficient for the left and right hemispheres. Physiological prior adjacency matrix matrix elements, These are unnormalized prior connection weights.

[0025] Furthermore, dynamic adjacency matrix The dimension is , The number of channels representing brainwave signals. The elements represent the channel connection weights learned through training data.

[0026] Furthermore, in step 2, the hybrid adjacency matrix... ,in, This is the fusion coefficient.

[0027] Furthermore, step 3 specifically includes:

[0028] Define the spatial-spectral feature sequence as ,in, Indicates the number of time windows. This represents the spatial-spectral feature dimension corresponding to each time window;

[0029] Set up a learnable sentiment anchor matrix: ,in, The number of emotional anchors;

[0030] Will Along the spatial-spectral feature dimension Cut evenly into The first subspace obtained after the nth subspace Subspace features , among which, time , ;

[0031] Each emotional anchor point Along the spatial-spectral feature dimension Cut evenly into The first subspace obtained after the nth subspace Subspace Anchor Point ;

[0032] Calculation via nonlinear mapping and Timing response matching degree between ;

[0033] Based on the set gate scalar Perform aggregation to obtain the first... Subspace, the first An aggregate vector of emotional anchor points: ;

[0034] All After bitwise fusion, the global sentiment representation is obtained through flattening and fully connected mapping.

[0035] Furthermore, gated scalar Specifically set as follows:

[0036]

[0037] in, It is the Sigmoid activation function. and The first Each subspace corresponds to a learnable gated weight vector and a bias term.

[0038] Furthermore, in step 4, the category prototype of the teacher network is the full-channel category prototype. The student network's category prototype is the few-channel category prototype. They are respectively:

[0039]

[0040]

[0041] in, This indicates that the current training batch belongs to the sentiment category. The set of indices of all samples, , These are the full-channel global sentiment representation and the few-channel global sentiment representation of the i-th sample, respectively, obtained through the sentiment anchor aggregation module.

[0042] Furthermore, in step 4, the teacher classification loss, student classification loss, prototype alignment loss, and knowledge distillation loss are respectively:

[0043]

[0044]

[0045]

[0046]

[0047] in, , , , These are teacher classification loss, student classification loss, prototype alignment loss, and knowledge distillation loss, respectively. The c-th element in the one-hot encoded vector of the true sentiment label; C represents the number of categories. Represents the probability of teacher classification The c-th element (obtained through the teacher network, the emotional anchor aggregation module, and the classifier); Represents the probability of student classification The c-th element (obtained through student network, sentiment anchor aggregation module, and classifier); This represents a category prototype for teacher networks. Represents the category prototype of the student network; This represents the set of valid sentiment categories present in the current training batch; Indicates the distillation temperature. Let KL divergence be a metric. and Let represent the original class score vectors output by the classifier for the teacher network and the student network, respectively, i.e., logits before Softmax normalization.

[0048] In another aspect, this invention provides a few-channel EEG emotion recognition system based on physiological prior guidance and cross-channel knowledge transfer, comprising:

[0049] The data preprocessing module is used to filter, downsample, segment, and extract differential entropy features from the input raw multi-channel EEG signals, and to construct paired EEG feature samples with full channels and few channels.

[0050] The multi-resolution channel configuration module is used to extract the spatial-spectral feature sequence of EEG features. It constructs a physiological prior adjacency matrix based on the distance relationship between EEG electrodes in the scalp space and the symmetrical electrode relationship between the left and right hemispheres. ; and the physiological prior adjacency matrix With the set learnable dynamic adjacency matrix The fusion yields a hybrid adjacency matrix. Then based on the hybrid adjacency matrix We construct normalized and scaled Laplacian matrices, and extract spatial-spectral feature sequences of EEG features through Chebyshev plot convolution.

[0051] The temporal feature aggregation module is equipped with an emotion anchor-based aggregation module. Based on the spatial-spectral feature sequence output by the multi-resolution channel configuration module, a global emotion representation is obtained through learnable emotion anchors, subspace attention matching, gating adjustment, and residual weighted aggregation. The global emotion representation is then input into a classifier to output the emotion category probability.

[0052] The cross-channel knowledge transfer module is used to perform alternating joint training between the full-channel teacher network and the few-channel student network;

[0053] The full-channel teacher network involves inputting full-channel EEG features into a multi-resolution channel configuration module to obtain full-channel spatial-spectral features. The few-channel student network involves inputting few-channel EEG features into a multi-resolution channel configuration module to obtain few-channel spatial-spectral features. ;Will The data is fed into the temporal feature aggregation module to obtain a global sentiment representation across all channels. Then Feed it into the classifier to obtain the teacher classification probability. ; and will The data is fed into the temporal feature aggregation module to obtain a global sentiment representation across all channels. Then Feed it into the classifier to obtain the student classification probability. ;

[0054] Furthermore, the joint losses used in alternating joint training include: teacher classification loss, student classification loss, prototype alignment loss, and knowledge distillation loss.

[0055] The emotion classification decision module, based on a trained few-channel student network, inputs the few-channel EEG features to be identified into the trained few-channel student network. After processing by the emotion anchor aggregation module and classifier, it outputs the probability of the emotion category and determines the final emotion recognition result based on the category with the highest probability.

[0056] The technical solution provided by this invention brings at least the following beneficial effects:

[0057] This invention addresses the challenges of spatial topological information loss, cross-channel distribution shift, and the difficulty in deploying full-channel models in few-channel EEG emotion recognition. It achieves highly reliable emotion recognition under few-channel conditions by constructing paired EEG feature samples between full-channel and few-channel systems, a physiological prior map convolutional module, an emotion anchor aggregation module, and a cross-channel knowledge transfer training mechanism between the full-channel teacher network and the few-channel student network. Specifically, the physiological prior map convolutional module integrates electrode spatial distance, left-right hemisphere symmetry, and data-driven dynamic connectivity to compensate for the spatial topological loss in few-channel EEG. The emotion anchor aggregation module enhances the dynamic representation of EEG emotions through learnable emotion anchors, subspace attention matching, gating modulation, and residual weighted aggregation. The cross-channel knowledge transfer training mechanism uses the global emotion representation, category prototypes, and soft labels of the full-channel teacher network to supervise and constrain the few-channel student network, enabling it to achieve near-full-channel model discrimination capabilities with a reduced number of electrodes. Therefore, this invention reduces the deployment cost and wearing complexity of few-channel EEG devices, improves the accuracy, robustness, and generalization ability of few-channel EEG emotion recognition, and is applicable to portable EEG emotion recognition scenarios. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 This is a flowchart of the overall method in an embodiment of the present invention.

[0060] Figure 2 This is a diagram of the basic feature extraction network framework in an embodiment of the present invention.

[0061] Figure 3 This is a diagram of the cross-channel knowledge transfer training framework based on ATSJT in an embodiment of the present invention.

[0062] Figure 4 The figure shows the performance curves of the experimental results in the embodiments of the present invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be described in detail and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Generally, the components of the embodiments of the present invention described and shown in the accompanying drawings can be arranged and designed using different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of the present invention.

[0064] This invention proposes a few-channel EEG emotion recognition method based on physiological prior guidance and cross-channel knowledge transfer, which includes an offline training phase and an online inference phase, such as... Figure 1 As shown.

[0065] During the offline training phase, full-channel and few-channel EEG signals corresponding to the same EEG sample are acquired and obtained after undergoing the same or corresponding preprocessing procedures to obtain full-channel EEG features. and few-channel EEG characteristics Inputting full-channel EEG characteristics into the full-channel teacher network. Small-channel EEG feature input, small-channel student network The full-channel teacher network is used to learn global sentiment semantics under the condition of complete spatial topology, while the few-channel student network is used to learn sentiment discrimination representations that can still be deployed under the condition of reduced electrode number. During training, constraints such as classification supervision, category prototype alignment, feature alignment, or knowledge distillation are used to make the output representation or output probability of the few-channel student network close to that of the full-channel teacher network.

[0066] During the online inference phase, only the trained, low-channel student network is retained. In actual deployment, the low-channel EEG signals collected by the low-channel EEG acquisition device are sent to the low-channel student network. After preprocessing, spatial feature extraction, temporal feature aggregation, and classification decision, the emotion classification result is output. The full-channel EEG acquisition device is no longer needed, nor is the full-channel teacher network required.

[0067] In one embodiment, the specific implementation steps of the few-channel EEG emotion recognition method based on physiological prior guidance and cross-channel knowledge transfer provided in this embodiment include:

[0068] Step 1: Construct paired EEG feature samples with full-channel and few-channel configurations.

[0069] The raw multichannel EEG signals are obtained from publicly available EEG emotion datasets (such as the SEED and SEED-IV datasets), and the raw multichannel EEG signals are filtered, power frequency interference suppressed, downsampled, segmented by sliding window, and frequency domain features extracted to obtain the EEG feature tensor.

[0070] Furthermore, the complete 62-channel features from the same EEG sample were retained as full-channel EEG features, and corresponding electrode channels were extracted according to the target few-channel configuration to obtain few-channel EEG features. This ensured a one-to-one correspondence between the full-channel and few-channel EEG features in terms of subject identity, experimental session, stimulus segment, time window, and emotion label, forming paired samples. Subsequently, the paired samples were divided into training and test sets in an 8:2 ratio. 80% of the training set data was used for the joint optimization of the full-channel teacher network and the few-channel student network, as well as for training the feature aggregation module. The remaining 20% ​​of the test set data was used as unknown samples and only input into the trained few-channel student network to rigorously verify the model's real emotion recognition performance under low-density channel configuration.

[0071] Step 2: Extract spatial-spectral feature sequences based on the physiological prior graph convolution module.

[0072] To alleviate the loss of spatial topological information caused by the reduced number of EEG electrodes, a physiological prior adjacency matrix was constructed based on the distance relationship between EEG electrodes in the scalp space and the symmetrical electrode relationship between the left and right hemispheres. At the same time, a learnable dynamic adjacency matrix was set to capture data-driven channel connectivity relationships. The physiological prior adjacency matrix and the dynamic adjacency matrix were fused to obtain a hybrid adjacency matrix. Based on the hybrid adjacency matrix, a normalized Laplacian matrix and a scaled Laplacian matrix were constructed, and the spatial-spectral feature sequence of the EEG signal was extracted by Chebyshev diagram convolution.

[0073] Step 3: Extract global sentiment representation based on the sentiment anchor aggregation module.

[0074] To address the dynamic changes in EEG emotional states over time, an emotional anchor aggregation module is implemented in both the full-channel teacher network and the few-channel student network. The spatial-spectral feature sequence is input into the emotional anchor aggregation module. Through learnable emotional anchors, subspace attention matching, gating modulation, and residual weighted aggregation, a global emotional representation is obtained, and the probability of the emotional category is output based on this global emotional representation.

[0075] Step 4: Conduct cross-channel knowledge transfer training based on the full-channel teacher network and the few-channel student network.

[0076] A full-channel teacher network and a few-channel student network are constructed. The full-channel teacher network receives full-channel EEG features to learn global sentiment semantics under complete spatial topology conditions; the few-channel student network receives few-channel EEG features to learn few-channel sentiment representations under deployable conditions. During training, the global sentiment representations, category prototypes, or soft labels output by the teacher network are used to supervise and constrain the student network. Combined with teacher classification loss, student classification loss, prototype alignment loss, and knowledge distillation loss, cross-channel knowledge transfer training is completed, enabling the few-channel student network to achieve sentiment discrimination capabilities close to those of the full-channel teacher network with a reduced number of electrodes.

[0077] Step 5: Perform online inference based on the trained few-channel student network.

[0078] During the online inference phase, only the trained few-channel student network is retained. Based on test samples from a publicly available EEG emotion dataset, corresponding few-channel EEG signals are extracted according to a preset few-channel configuration, and a preprocessing procedure consistent with or corresponding to the training phase is performed to obtain few-channel EEG features. These few-channel EEG features are then input into the trained few-channel student network, processed by a physiological prior graph convolution module, an emotion anchor aggregation module, and a classifier, to output emotion category probabilities. The final emotion recognition result is determined based on the category with the highest probability.

[0079] In one embodiment, step 1 includes the following steps:

[0080] Step 1-1: Obtain the raw multi-channel EEG signals.

[0081] The raw multichannel EEG signals can come from a full-channel (e.g., 62-channel) EEG device, or from a full-channel EEG dataset that you have collected yourself.

[0082] Steps 1-2, after preprocessing, yield the EEG feature tensor:

[0083]

[0084] in, Indicates the number of time windows. This indicates the number of channels currently receiving EEG signals, for example... , This represents the dimension of the frequency domain features extracted for each channel; when using the differential entropy features of five standard frequency bands, .

[0085] By preserving the complete set of channels from the same EEG sample, we can obtain full-channel EEG features:

[0086]

[0087] Steps 1-3 preserve the complete channel set from the same EEG sample to obtain full-channel EEG features. Based on the target few-channel configuration (e.g., 32, 16, 8, 4 channels, etc.), corresponding electrode channels are extracted to obtain few-channel EEG characteristics. Based on the target few-channel configuration, corresponding electrode channels are extracted to obtain few-channel EEG characteristics:

[0088]

[0089] in, Indicates the total number of channels. This indicates a small number of channels, and typically satisfies... .

[0090] Steps 1-4, make and Matched training samples are formed by maintaining correspondences in subject identity, stimulus fragments, time windows, and emotional labels. This pairing relationship enables the teacher network and student network to align their representations under the same emotional sample.

[0091] In a preferred embodiment, the few-channel configuration includes at least one of 32 channels, 16 channels, 8 channels, and 4 channels. Specifically, 32 channels are used to simulate medium-density EEG devices, 16 channels are used to simulate high-level portable EEG devices, and 8 and 4 channels are used to simulate minimalist portable EEG devices. Different channel configurations can be used to train corresponding student networks separately, or a unified student network can be trained by sharing backbone parameters and setting channel masks.

[0092] In one embodiment, see Figure 2 Step 2 includes the following steps:

[0093] To alleviate the loss of spatial topological information caused by the reduced number of electrodes (32, 16, 8, 4 channels), this invention incorporates a physiological prior graph convolutional module (PG-DGCNN) in both the teacher and student networks. This module comprises four sub-steps: physiological prior adjacency matrix construction, dynamic adjacency matrix learning, hybrid adjacency matrix fusion, and graph convolutional spatial feature extraction.

[0094] Step 2-1: Calculate the distance between the electrodes based on their positions in the scalp space. Simultaneously, the hemispherical symmetrical indication quantity is determined based on the symmetrical electrode pairs of the left and right hemispheres. For the first The first electrode and the second For each electrode, its unnormalized prior connection weights can be expressed as:

[0095]

[0096] in, For distance scale parameters, The coefficient for enhancing symmetry between the left and right hemispheres.

[0097] Step 2-2, for Normalization is performed to obtain the physiological prior adjacency matrix. Its matrix elements are represented as:

[0098]

[0099] This physiological prior adjacency matrix contains both local spatial proximity relationships and cross-hemispherical symmetry relationships, which can provide stable structural constraints for graph convolution when there are few channel nodes.

[0100] Steps 2-3: Set up the dynamic adjacency matrix The dynamic adjacency matrix is ​​a learnable parameter matrix. Its elements represent the channel connection weights learned through training data. To ensure training stability, It can be processed by Sigmoid, Softmax, or nonnegative normalization.

[0101] Steps 2-4, will and Merge into a hybrid adjacency matrix:

[0102]

[0103] in, The fusion coefficient can be set as a learnable parameter or preset according to experiments. When the size is small, the model relies more on physiological priors; when When the size is large, the model relies more on data-driven dynamic connections.

[0104] Steps 2-5, based on the hybrid adjacency matrix We construct the degree matrix, the normalized Laplacian matrix, and the scaled Laplacian matrix, and use Chebyshev diagram convolution to extract spatial-spectral feature sequences.

[0105] Specifically, based on the hybrid adjacency matrix Construct the degree matrix :

[0106]

[0107] in, This indicates the number of channels currently receiving EEG signals. Indicates the first The brainwave channel and the first Hybrid connectivity weights between individual EEG channels Indicates the first The degree value corresponding to each EEG channel.

[0108] In one embodiment, based on the degree matrix and hybrid adjacency matrix Constructing the normalized Laplace matrix :

[0109]

[0110] in, express An identity matrix of order 1. Degree matrix A negative 1 / 2 power matrix.

[0111] To reduce the computational overhead of eigenvalue decomposition during graph convolution and to ensure that the eigenvalues ​​of the Laplacian matrix fall within the stable approximation interval of the Chebyshev polynomial, the normalized Laplacian matrix... After scaling, the scaled Laplacian matrix is ​​obtained. :

[0112]

[0113] in, Represents the normalized Laplace matrix The largest eigenvalue, This represents the scaled Laplacian matrix. Through the above scaling process, the eigenvalues ​​of the Laplacian matrix can be constrained to... The interval is determined so that a stable approximation can be made using Chebyshev polynomials.

[0114] Subsequently, based on the scaled Laplacian matrix implement Chebyshev graph convolution yields spatial-spectral feature sequences. :

[0115]

[0116] in, Denotes the order of the Chebyshev polynomial. Indicates the first Chebyshev basis functions of order, This represents the input EEG feature tensor. and They represent the first The learnable weight matrix and bias terms corresponding to graph convolution of order 1. This represents the spatial-spectral feature sequence of the output. This corresponds to a multilayer perceptron (MLP) or a fully connected mapping layer.

[0117] Alternatively, the Chebyshev basis functions are calculated recursively as follows:

[0118]

[0119]

[0120]

[0121] in, This represents the initial Chebyshev basis function value. This represents the value of the first-order Chebyshev basis function.

[0122] Through the aforementioned spatial modeling process, the few-channel student network can learn more stable brain region association features by utilizing physiological prior constraints and dynamic graph structures under conditions with a smaller number of electrodes, thereby alleviating the problem of missing spatial topological information in few-channel EEG signals.

[0123] The full-channel teacher network and the few-channel student network each maintain independent hybrid adjacency matrices. and .

[0124] In one embodiment, see Figure 3 Step 3 includes the following steps:

[0125] EEG emotional states exhibit distinct dynamic evolutionary characteristics. To avoid overlooking short-term emotional microstates using ordinary pooling or simple recurrent structures, this invention incorporates an emotional anchor aggregation module in both the teacher and student networks.

[0126] Step 3-1, denote the spatial-spectral feature sequence output by the graph convolution module as:

[0127]

[0128] in, Indicates the number of time windows. This represents the spatial-spectral feature dimension corresponding to each time window after processing by the PG-DGCNN graph convolution module and feature mapping layer. Not equivalent to input feature dimension , The original frequency domain feature dimension for each channel, and This refers to the higher-order feature dimension output by the network, the value of which can be preset according to the network structure.

[0129] Step 3-2, Set up a learnable emotional anchor matrix:

[0130]

[0131] in, The emotional anchor matrix E represents the number of emotional anchors. These anchors are not required to equal the number of emotional categories or the number of EEG channels; rather, they are used to represent several learnable emotional reference patterns in the latent space. The emotional anchor matrix E is initialized before model training begins. As one implementation method, random initialization can be used.

[0132] Step 3-3, specifically, firstly, from the spatial-spectral feature sequence Extract the first Spatial-spectral feature vectors corresponding to each time window ,in, Subsequently, and every emotional anchor It is uniformly divided into M independent subspaces (Heads) along the feature dimension, that is... , Indicates along the feature dimension Cut evenly into The first subspace obtained after the nth subspace Subspace features, This represents the anchor point after segmentation. Within the m-th subspace, the input feature frame is calculated using a nonlinear mapping. With emotional anchors The temporal response matching degree between them (i.e., attention weights) :

[0133]

[0134] in, Representation of spatial-spectral feature sequences The Middle Feature slices corresponding to each time window; This indicates the number of global sentiment anchors. Indicates the first An emotional anchor index, Indicates the first Subspace indexes. Indicates the first The corresponding subspace within the first A learnable weight vector of sentiment anchors is used to perform weighted matching of each dimension of the current subspace features; This represents the corresponding learnable bias term, used to adjust the base response strength of this emotional anchor. The numerator term represents the subspace features of the current time frame on the basis of the first... The nonnormalized response intensity of each emotional anchor point; the denominator term represents the subspatial features of the current time frame on all The sum of the response intensities of each emotional anchor is used for Softmax normalization. The final result is... Indicates the feature of the current time frame in the th order. In each subspace, belonging to the first The soft weighting of each emotional anchor point.

[0135] Steps 3-4: Set the gated scalar Used for dynamic adjustment of the first The time frame in the ... The contribution of each subspace to the final aggregation result. Aggregation based on residual form:

[0136]

[0137] in, Indicates the first Subspace, the first An aggregated vector of emotional anchor points Indicates that for the first A dynamic gating scalar for each subspace, used to measure time. The degree to which the spatial-spectral features contribute to the final sentiment representation aggregation within this subspace. This gating scalar can be determined by the current subspace features. It is obtained through a learnable linear mapping and the Sigmoid function, i.e.

[0138]

[0139] in, This refers to the Sigmoid activation function. and The first Each subspace corresponds to a learnable gated weight vector and a bias term. Through this gating term, the model can adaptively enhance features in subspaces that are more important for sentiment discrimination and suppress feature groups that contribute less or are noisier.

[0140] Steps 3-5 involve merging the aggregated vectors corresponding to all subspaces and all sentiment anchors bit by bit, and then flattening and fully connected mapping to obtain the global sentiment representation. ;

[0141] Global sentiment representation Input a classifier, and the classifier outputs the probability of the sentiment category. The classifier can be a Softmax classifier.

[0142] This approach enhances the model's ability to express emotional fluctuations at different time points and reduces the impact of changes in the number of channels on the global representation.

[0143] In one embodiment, step 4 includes the following steps:

[0144] In this invention, cross-channel knowledge transfer training is used to transfer the global sentiment semantics of the full-channel teacher network to at least the channel student network. This training process can employ alternating teacher-student joint training or a joint training method with teacher forward supervision and student backward updating. Due to the different input channel dimensions, the teacher network and student network each have independent spatial feature extraction modules (i.e., PG-DGCNN), denoted as follows: and ,in This refers to the global sentiment representation module proposed in step 3, which is based on the sentiment anchor aggregation module. This represents the Softmax classifier.

[0145] Step 4-1, Online training for teachers. This involves analyzing the full-channel EEG characteristics. Input the teacher network to obtain the full-channel spatial-spectral characteristics. Full-channel global sentiment representation Teacher classification probability Calculating Teacher Classification Loss Based on Real Sentiment Labels It is used to update the teacher network and shareable semantic aggregation parameters.

[0146] During the teacher training phase, calculations were performed based on full-channel EEG samples:

[0147]

[0148] The teacher network classification loss is:

[0149]

[0150] in, The c-th element in the one-hot encoded vector representing the true sentiment label.

[0151] Step 4-2, Student online training. This involves using limited-channel EEG features. Input the student network to obtain the few-channel spatial-spectral characteristics. Few-channel global sentiment representation Student classification probability Calculating few-channel classification loss based on real sentiment labels. .

[0152] During the student phase, calculations were performed based on a limited number of EEG samples:

[0153]

[0154] The student network classification loss is:

[0155]

[0156] Step 4-3, for each valid sentiment category in the current training batch Calculate the full-channel category prototype and the few-channel category prototype separately:

[0157]

[0158] in, This indicates that the current training batch belongs to the sentiment category. The set of indices of all samples. Then, the prototype alignment loss is calculated:

[0159]

[0160] in, This represents the set of valid sentiment categories present in the current training batch.

[0161] Step 4-4, Knowledge Distillation Constraints. Based on the classification logits output by the teacher and student networks, calculate the knowledge distillation loss:

[0162]

[0163] in, and Let represent the original class score vectors output by the full-channel teacher network and the few-channel student network before Softmax normalization, respectively. Indicates the distillation temperature. Let KL divergence be denoted as KL divergence.

[0164] Steps 4-5: Total loss of the student network. The training objective for the student network can be defined as:

[0165]

[0166] in, and The loss weights are used. Through this training objective, the student network not only learns the sentiment classification boundary under limited-channel input, but also learns the category structure and output distribution formed by the teacher network under full-channel conditions.

[0167] In one embodiment, step 5 includes the following steps:

[0168] The online reasoning phase includes the following steps:

[0169] Step 5-1: Use a few-channel EEG acquisition device to acquire a few-channel EEG signals. The device can be a portable EEG device with 16 channels, 8 channels, or 4 channels.

[0170] Step 5-2: Perform the same or corresponding preprocessing procedure as the training phase on the few-channel EEG signals to obtain few-channel EEG features. .

[0171] Step 5-3, The trained few-channel student network is input, and spatial-spectral features are obtained through the physiological prior graph convolution module. Global sentiment representation is obtained through the sentiment anchor aggregation module, and the sentiment category probability is output through the classifier.

[0172] Step 5-4: Take the category with the highest probability as the emotion recognition result, or output multiple emotion category probabilities for subsequent warnings, human-computer interaction, or psychological state assessment.

[0173] Figure 4 The emotion recognition accuracy of the embodiments of this invention and the baseline method on the SEED and SEED-IV datasets is presented under different numbers of EEG channels. Figure 4 It can be seen that the recognition accuracy of the baseline method decreases on both datasets as the number of EEG channels decreases. Specifically, on the SEED dataset, the accuracy decrease of the baseline method is relatively gradual in the range of 62 to 16 channels, but the accuracy decrease becomes significantly larger when the number of channels is further reduced to 8 and 4 channels. On the SEED-IV dataset, the accuracy of the baseline method shows a more pronounced and continuous downward trend with the decrease in the number of channels. When the number of channels is reduced from 32 to 16, the accuracy decreases by nearly 10%, indicating that the lack of spatial topological information under fewer channels has a significant impact on emotion recognition performance. Compared to the baseline method, the embodiments of this invention achieve higher recognition accuracy under different fewer-channel configurations. Specifically, with an 8-channel configuration on the SEED dataset, the accuracy improvement reaches 4.35%; with 32 to 8-channel configurations on the SEED-IV dataset, the accuracy improvement is approximately 2.3% to 3.6%. Experimental results show that this invention, through physiological prior map convolution, emotional anchor aggregation, and cross-channel knowledge transfer between the full-channel teacher network and the few-channel student network, can compensate for some missing spatial topology and emotional semantic information when the number of EEG channels is reduced, thereby improving the accuracy and robustness of few-channel EEG emotion recognition. Considering both recognition performance and device deployment cost, the 16-channel configuration can maintain recognition accuracy of over 93% and over 82% on the SEED and SEED-IV datasets, respectively, while reducing the number of electrodes and wearing complexity, demonstrating good value for the deployment of portable EEG devices.

[0174] In one embodiment, this embodiment also provides a few-channel EEG emotion recognition system based on physiological prior guidance and cross-channel knowledge transfer for implementing the few-channel EEG emotion recognition method provided in this embodiment. Based on a hierarchical design principle, it consists of a data preprocessing layer, a multi-resolution channel configuration layer, a spatial feature extraction layer, a temporal feature aggregation layer, a cross-channel knowledge transfer layer, and an emotion classification decision layer. The functions of each layer are as follows:

[0175] Data Preprocessing Layer: This layer consists of four functional modules: signal filtering, downsampling, temporal segmentation, and differential entropy feature extraction. It primarily provides EEG feature sample data in a specified format for subsequent model algorithms. Specifically, the signal filtering module performs bandpass filtering on the raw EEG signal to remove power frequency interference and high-frequency noise; the downsampling module adjusts the signal sampling rate to a uniform standard; the temporal segmentation module uses a fixed-length overlapping sliding window technique to segment the continuous EEG sequence into sample tensors of fixed size; and the differential entropy feature extraction module calculates differential entropy features from five standard frequency bands as the basic input.

[0176] Multi-resolution channel configuration layer: This layer dynamically extracts 32, 16, 8, or 4 channel subsets from the full-channel data based on the hardware constraints of the target application scenario, and generates strictly aligned paired datasets. This layer incorporates core channel subset configuration rules based on the electrode distribution statistics of mainstream commercial devices, ensuring that spatial information related to sentiment discrimination is preserved to the maximum extent at different resolutions.

[0177] Spatial Feature Extraction Layer: This layer deploys a physiologically prior-guided dynamic graph convolutional network (PG-DGCNN) to model the spatial relationships between different EEG channels. This module uses a hybrid adjacency matrix that combines prior and adaptive approaches to explicitly inject the theory of hemispherical asymmetry induced by emotions into the graph structure construction while preserving data-driven learning capabilities, providing strong regularization constraints for scenarios with few channels.

[0178] Temporal Feature Aggregation Layer: This layer deploys an Emotional Anchor-Based Feature Aggregation Network (EAFA) to extract global emotional representations that are insensitive to the number of channels and have strong generalization capabilities in the temporal dimension. This module models the dynamic evolution of emotion as a cross-attentional interaction process between temporal feature frames and implicit emotional anchors through a multi-head subspace mechanism, dynamically capturing the temporal fluctuation patterns of EEG sequences under different emotional microstates.

[0179] Cross-channel knowledge transfer layer: This layer is used to perform alternating joint training of the full-channel teacher network and the few-channel student network. Through a hard parameter sharing mechanism, the few-channel student network can reuse the high-order temporal and semantic fusion rules of the full-channel teacher network; through fine-grained sentiment prototype alignment loss, the few-channel network is forced to move closer to the same prototype of the full-channel features in the feature space, thereby completing the missing global spatial topological semantics.

[0180] Sentiment Classification Decision Layer: This layer consists of a Softmax classifier and a decision output module. It outputs the probability distribution of sentiment states based on the aggregated global features and provides the final classification result. This layer outputs the probability distribution of sentiment states based on the aggregated global features and determines the final sentiment category based on the maximum probability.

[0181] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the foregoing embodiments have described the present invention in detail, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the embodiments of the present invention.

[0182] The above are merely some embodiments of the present invention. For those skilled in the art, various modifications and improvements can be made without departing from the inventive concept of the present invention, and all such modifications and improvements fall within the scope of protection of the present invention.

Claims

1. A few-channel EEG emotion recognition method based on physiological prior and cross-channel knowledge transfer, characterized in that, Includes the following steps: Step 1: Construct paired EEG feature samples with full-channel and few-channel configurations; Based on the full-channel EEG features of the input, the corresponding electrode channels are extracted according to the target few-channel configuration to obtain few-channel EEG features; The full-channel EEG features and the few-channel EEG features are kept in a one-to-one correspondence in terms of subject identity, experimental session, stimulus segment, time window and emotional label, forming paired EEG feature samples. Step 2: Extract spatial-spectral feature sequences based on the physiological prior map convolution module; A physiological prior adjacency matrix was constructed based on the distance relationship between EEG electrodes in the scalp space and the symmetrical relationship between electrodes in the left and right hemispheres. ; Configure a learnable dynamic adjacency matrix To capture data-driven channel connectivity; and to matrix and The fusion yields a hybrid adjacency matrix. ; Based on hybrid adjacency matrix We construct normalized and scaled Laplacian matrices and extract spatial-spectral feature sequences of EEG signals through Chebyshev plot convolution. Step 3: Extract global sentiment representation based on the sentiment anchor aggregation module; An emotion anchor aggregation module is set up in both the full-channel teacher network and the few-channel student network; the full-channel teacher network receives full-channel EEG features, and extracts the full-channel spatial-spectral feature sequence through the physiological prior map convolution module. The student network receives a small number of EEG features, which are then extracted using a physiological prior graph convolution module to obtain a small number of spatial-spectral feature sequences. ; The spatial-spectral feature sequence is input into the sentiment anchor aggregation module. Through learnable sentiment anchors, subspace attention matching, gating adjustment, and residual weighted aggregation, a global sentiment representation is obtained. The global sentiment representation is then input into a classifier to output the sentiment category probability. Step 4: Conduct cross-channel knowledge transfer training based on the full-channel teacher network and the few-channel student network; Full-channel global sentiment representation obtained using teacher network and sentiment anchor aggregation module The category prototype or soft label supervises and constrains the student network, and combines teacher classification loss, student classification loss, prototype alignment loss and knowledge distillation loss to achieve cross-channel knowledge transfer training. Step 5: Perform online inference based on the trained few-channel student network; During the online inference phase, the few-channel EEG features to be identified are input into the trained few-channel student network. After processing by the emotion anchor aggregation module and classifier, the probability of the emotion category is output, and the final emotion recognition result is determined based on the category with the highest probability.

2. The method as described in claim 1, characterized in that, In step 1, the tensor dimension of the full-channel EEG features / few-channel EEG features is: ,in, Indicates the number of time windows. The number of channels representing brainwave signals. This represents the dimension of the frequency domain features extracted for each channel.

3. The method as described in claim 1, characterized in that, In step 2, the physiological prior adjacency matrix for: in, For the first The first electrode and the second The distance between the electrodes For the first The first electrode and the second The hemispherical symmetry of each electrode indicates the quantity. For distance scale parameters, The enhancement coefficient is the symmetry enhancement factor between the left and right hemispheres. Physiological prior adjacency matrix matrix elements, These are unnormalized prior connection weights.

4. The method as described in claim 3, characterized in that, Dynamic adjacency matrix The dimension is , The number of channels representing brainwave signals. The elements represent the channel connection weights learned through training data.

5. The method as described in claim 4, characterized in that, In step 2, the hybrid adjacency matrix ,in, This is the fusion coefficient.

6. The method as described in claim 1, characterized in that, Step 3 specifically includes: Define the spatial-spectral feature sequence as ,in, Indicates the number of time windows. This represents the spatial-spectral feature dimension corresponding to each time window; Set up a learnable sentiment anchor matrix: ,in, The number of emotional anchors; Will Along the spatial-spectral feature dimension Cut evenly into The first subspace obtained after the nth subspace Subspace features , among which, time , ; Each emotional anchor point Along the spatial-spectral feature dimension Cut evenly into The first subspace obtained after the nth subspace Subspace Anchor Point ; Calculation via nonlinear mapping and Timing response matching degree between ; Based on the set gate scalar Perform aggregation to obtain the first... Subspace, the first An aggregate vector of sentiment anchors: ; All After bitwise fusion, the global sentiment representation is obtained through flattening and fully connected mapping.

7. The method as described in claim 6, characterized in that, Gated scalar Specifically set as follows: in, It is the Sigmoid activation function. and The first Each subspace corresponds to a learnable gated weight vector and a bias term.

8. The method as described in claim 1, characterized in that, In step 4, the category prototype of the teacher network is the full-channel category prototype. The student network's category prototype is the few-channel category prototype. They are respectively: in, This indicates that the current training batch belongs to the sentiment category. The set of indices of all samples, , These are the full-channel global sentiment representation and the few-channel global sentiment representation of the i-th sample, respectively, obtained through the sentiment anchor aggregation module.

9. The method as described in claim 8, characterized in that, In step 4, the teacher classification loss, student classification loss, prototype alignment loss, and knowledge distillation loss are respectively: in, , , , These are teacher classification loss, student classification loss, prototype alignment loss, and knowledge distillation loss, respectively. The c-th element in the one-hot encoded vector of the true sentiment label; C represents the number of categories. Represents the probability of teacher classification The c-th element in; Represents the probability of student classification The c-th element in; This represents a category prototype for teacher networks. This represents the category prototype of the student network; This represents the set of valid sentiment categories present in the current training batch; Indicates the distillation temperature. Let KL divergence be a metric. and These represent the original category score vectors output by the classifier for the teacher network and the student network, respectively.

10. A few-channel EEG emotion recognition system based on physiological prior knowledge and cross-channel knowledge transfer, characterized in that, include: The data preprocessing module is used to filter, downsample, segment, and extract differential entropy features from the input raw multi-channel EEG signals, and to construct paired EEG feature samples with full channels and few channels. The multi-resolution channel configuration module is used to extract the spatial-spectral feature sequence of EEG features. It constructs a physiological prior adjacency matrix based on the distance relationship between EEG electrodes in the scalp space and the symmetrical electrode relationship between the left and right hemispheres. ; and physiological prior adjacency matrix With the set learnable dynamic adjacency matrix The fusion yields a hybrid adjacency matrix. ; Then based on the hybrid adjacency matrix We construct normalized and scaled Laplacian matrices, and extract spatial-spectral feature sequences of EEG features through Chebyshev plot convolution. The temporal feature aggregation module is equipped with an emotion anchor-based aggregation module. Based on the spatial-spectral feature sequence output by the multi-resolution channel configuration module, a global emotion representation is obtained through learnable emotion anchors, subspace attention matching, gating adjustment, and residual weighted aggregation. The global emotion representation is then input into a classifier to output the emotion category probability. The cross-channel knowledge transfer module is used to perform alternating joint training between the full-channel teacher network and the few-channel student network; The full-channel teacher network involves inputting full-channel EEG features into a multi-resolution channel configuration module to obtain full-channel spatial-spectral features. The few-channel student network involves inputting few-channel EEG features into a multi-resolution channel configuration module to obtain few-channel spatial-spectral features. ;Will The data is fed into the temporal feature aggregation module to obtain a global sentiment representation across all channels. Then Feed it into the classifier to obtain the teacher classification probability. ; and will The data is fed into the temporal feature aggregation module to obtain a global sentiment representation across all channels. Then Feed it into the classifier to obtain the student classification probability. ; Furthermore, the joint losses used in alternating joint training include: teacher classification loss, student classification loss, prototype alignment loss, and knowledge distillation loss. The emotion classification decision module, based on a trained few-channel student network, inputs the few-channel EEG features to be identified into the trained few-channel student network. After processing by the emotion anchor aggregation module and classifier, it outputs the probability of the emotion category and determines the final emotion recognition result based on the category with the highest probability.

Images