Cross-subject domain adaptive SSVEP electroencephalogram classification method based on self-training

Abstract

The invention relates to a self-training-based cross-subject-domain adaptive SSVEP electroencephalogram classification method, which comprises the following steps of: acquiring an SSVEP signal, and capturing multi-band information of the SSVEP signal to carry out filter bank European alignment; training a convolutional neural network model by using the aligned SSVEP signals; in the pre-training stage, the aligned SSVEP signals are input into a convolutional neural network model, target features are obtained, a domain discriminator is introduced for aligning feature distribution of different domains, meanwhile, a classification head is used for classifying the target features, and a pre-trained electroencephalogram classification model is obtained; in the self-training stage, a pre-trained electroencephalogram classification model is divided into a teacher model and a student model, the teacher model is utilized to predict target domain data, pseudo-tags are generated through confidence coefficient screening, a pseudo-tag fusion strategy and a time-frequency enhancement contrast learning strategy are further introduced to guide learning of the student model, and the learning efficiency of the student model is improved. And obtaining a final electroencephalogram classification model.

Description

Technical Field

[0001] This invention relates to the fields of brain-computer interface and machine learning technology, and in particular to a self-training-based cross-subject domain adaptive SSVEP EEG classification method. Background Technology

[0002] Brain-computer interface (BCI) technology establishes a direct communication pathway between the brain and external devices, enabling novel information interaction methods independent of the peripheral nervous system. Among various BCI paradigms, steady-state visual evoked potentials (SSVEPs) have attracted significant attention due to their high signal-to-noise ratio and user-friendliness. SSVEPs are rhythmic electrophysiological signals induced in the visual cortex of the brain when a user gazes at a periodically flashing visual stimulus at a specific frequency. These signals are synchronized with the stimulus frequency and have been widely applied in areas such as spelling input and device control. To achieve widespread applicability of BCI systems, cross-subject technology aims to enable the system to perform high-performance EEG decoding for new users without relying on or requiring only minimal user-specific data calibration.

[0003] However, existing SSVEP-based BCI technologies still face significant challenges in cross-subject applications, primarily in two aspects: First, inherent differences exist among individual subjects in brain structure and neural activity patterns, directly leading to significant inconsistencies in the distribution of key features of their SSVEP signals, such as amplitude, phase, and frequency response. Second, to achieve high-precision classification, existing methods often require collecting a large amount of labeled SSVEP data for each target subject for model calibration, a time-consuming and labor-intensive process that severely hinders practical deployment. Existing solutions, such as untrained CCA and FBCCA methods, exhibit poor generalization and low accuracy; domain-based generalization methods do not utilize target user data and have limited performance; while previous domain adaptation methods struggle to address the issues of low-quality pseudo-labels and error accumulation. Therefore, a new scheme capable of efficiently and accurately performing cross-subject EEG decoding is urgently needed. Summary of the Invention

[0004] To address the problems existing in the prior art, the purpose of this invention is to provide a self-trained, cross-subject domain adaptive SSVEP EEG classification method to overcome the technical shortcomings of existing SSVEP-BCI cross-subject classification methods, such as heavy calibration burden, large individual variability, and limited decoding accuracy. This invention can achieve high-performance cross-subject classification without relying on target subject labeled data.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A self-training-based cross-subject domain adaptive SSVEP EEG classification method includes:

[0007] Acquire the SSVEP signal, capture the multi-band information of the SSVEP signal, and perform Euclidean alignment of the filter bank.

[0008] A convolutional neural network model is trained using the aligned SSVEP signal; the training process includes a pre-training phase and a self-training phase.

[0009] In the pre-training stage, the aligned SSVEP signal is input into a convolutional neural network model to obtain target features. The target features are classified using a classification head, and a domain discriminator is introduced to align the feature distributions of different domains to obtain a pre-trained EEG classification model.

[0010] During the self-training phase, the pre-trained EEG classification model is divided into a teacher model and a student model. The teacher model is used to predict target domain data, and pseudo-labels are generated through confidence screening. Furthermore, a pseudo-label fusion strategy and a time-frequency enhanced contrastive learning strategy are introduced to guide the learning of the student model, thereby obtaining an EEG classification model. The EEG classification model is then used for SSVEP EEG classification.

[0011] Optionally, capturing the multi-band information of the SSVEP signal and performing Euclidean alignment of the filter bank includes:

[0012] ;

[0013] in, For the aligned signal, The mean covariance matrix for all samples is... This is the original signal.

[0014] Optionally, the convolutional neural network model is connected to the domain discriminator through a gradient inversion layer, which can invert the gradients flowing through the convolutional neural network model.

[0015] Optionally, the loss function in the pre-training phase includes:

[0016] ;

[0017] ;

[0018] ;

[0019] in, For source domain classifier, The cross-entropy loss is the source domain. For classification header, For feature extractors, For the true label of the source domain, For the domain discriminator loss, For the trainable parameters of the feature extractor G, Let D be the trainable parameters of the domain discriminator. For the discrimination loss of the source domain, For source domain feature extractor, For the discrimination loss of the target domain, For domain discriminators.

[0020] Optionally, obtaining the EEG classification model includes:

[0021] The teacher model is used to predict target domain data, and pseudo-labels are generated by confidence screening. The pseudo-label fusion strategy is further introduced to integrate multiple augmented views to guide the learning of the student model and obtain the EEG classification model.

[0022] While generating the pseudo-labels, the steady-state visual evoked potential signals under the pseudo-labels are subjected to temporal perturbation and noise injection along the time and spectral dimensions, respectively, to generate diversified views and enhance feature discrimination capabilities.

[0023] Optionally, integrating the multiple enhanced views includes:

[0024] ;

[0025] ;

[0026] in, Indicates projection features, Indicates projection relative to the original view Calculate the cosine similarity. The final prediction result after fusion of multiple enhanced views. To enhance the indexing of views, The weight coefficient for the k-th augmented view. This represents the model prediction result corresponding to the k-th augmented view. The projection feature vector obtained by passing the projector through the k-th augmented view. This is the projection feature vector corresponding to the original view. The projection feature vector obtained by passing the m-th enhanced view through the projector. This is a numerical stability term used to avoid the denominator being zero. This is the input sample for the k-th enhanced view.

[0027] Optionally, guiding the learning of the student model includes:

[0028] The parameters of the teacher model are updated using an exponential moving average of the student model parameters:

[0029] ;

[0030] in, For the parameters of the teacher model, For student model parameters, The momentum coefficient of the exponential moving average.

[0031] Optionally, the loss function in the self-training phase includes:

[0032] ;

[0033] ;

[0034] in, This is the loss function for the self-training phase. The loss of the student model for classifying samples in the target domain. The weight hyperparameters of the learning loss are used to adjust their proportion in the total loss. The contrastive learning loss is used with sample i as the anchor point. Let i be the set of positive samples. The sample index is the value in the set of positive samples. Let be the projected feature vector of sample i. This is the transpose of the vector. Let p be the projected feature vector of the positive sample p. For temperature coefficient, The sample index in the candidate sample set. Let i be the set of samples corresponding to anchor point sample i.

[0035] The beneficial effects of this invention are as follows:

[0036] This invention designs a filter bank Euclidean alignment strategy to fully utilize the frequency information in the SSVEP filter bank; secondly, it proposes a cross-subject self-training framework, which includes two stages: pre-training through domain adversarial learning to align the source and target domain distributions; and dual ensemble self-training to optimize pseudo-label quality; and introduces a time-frequency enhancement contrastive learning module to enhance feature discrimination capabilities across multiple enhanced views.

[0037] This invention targets steady-state visual evoked potential tasks. First, it decomposes multi-channel signals using filter banks. Then, it calculates the average covariance matrix based on all samples in the dimension of "number of channels × number of filter banks" and performs Euclidean alignment accordingly to better capture its harmonic components.

[0038] This invention combines adversarial learning to align domain features and improve pseudo-label quality during the pre-training stage. During the self-training stage, it optimizes pseudo-labels by integrating teacher-student models in a dual integration mechanism and fusing multi-view pseudo-labels, thereby reducing error accumulation.

[0039] This invention generates different augmented views by applying two time-domain and frequency-domain data augmentation operations—amplitude scaling and noise injection—to steady-state visual evoked potential signals. It then uses a supervised contrastive loss function to learn highly discriminative feature representations on the time-frequency augmented views. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the 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.

[0041] Figure 1 This is a schematic diagram of a self-training cross-subject domain adaptive SSVEP EEG classification method according to an embodiment of the present invention;

[0042] Figure 2 This is a schematic diagram of the Euclidean alignment of the filter bank according to an embodiment of the present invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] like Figure 1As shown, this embodiment discloses a self-training-based cross-subject domain adaptive SSVEP EEG classification method, including: acquiring SSVEP signals, capturing multi-band information of the SSVEP signals and performing Euclidean alignment of the filter bank; training a convolutional neural network model using the aligned SSVEP signals; the training process includes a pre-training stage and a self-training stage; in the pre-training stage, the aligned SSVEP signals are input into the convolutional neural network model to acquire target features, the target features are classified using a classification head, and a domain discriminator is introduced to align the feature distributions of different domains to obtain a pre-trained EEG classification model; in the self-training stage, the pre-trained EEG classification model is divided into a teacher model and a student model, the teacher model is used to predict target domain data, and pseudo-labels are generated through confidence filtering, and a pseudo-label fusion strategy and a time-frequency enhanced contrastive learning strategy are further introduced to guide the learning of the student model to obtain an EEG classification model, and the EEG classification model is used for SSVEP EEG classification.

[0046] Specifically, this embodiment discloses a self-training-based cross-subject domain adaptive SSVEP EEG classification method, including: firstly, global sample data alignment based on the characteristics of SSVEP signals to alleviate individual differences at the preprocessing level; then, introducing adversarial learning in the pre-training stage to reduce inter-domain distribution differences at the feature level; furthermore, designing a dual-integration self-training mechanism to optimize the quality of pseudo-labels in the target domain; and simultaneously, designing a time-frequency enhanced contrastive learning module to enhance the model's robustness to noise fluctuations. Through the systematic integration of the above technologies, this invention can achieve high-precision cross-subject SSVEP classification, providing effective support for the practical application and promotion of brain-computer interfaces.

[0047] The data alignment module is used in the preprocessing stage to propose a filter bank Euclidean alignment strategy specifically designed for SSVEP signals. This strategy is achieved by calculating the global average covariance matrix of the filter bank-channel joint space and performing a whitening transformation on the signal, thereby making better use of the complementary information of the signal at different frequencies.

[0048] A cross-subject self-training module is included, encompassing both pre-training and self-training. During pre-training, adversarial learning is employed to align the distributions of the source and target domains, thereby enhancing the model's generalization ability. In the self-training phase, a dual-ensemble self-training mechanism is proposed. Specifically, the time-ensembled teacher model generates multi-view ensemble pseudo-labels to improve pseudo-label quality.

[0049] To address the issue that self-training is still affected by noise labels, a time-frequency enhanced contrastive learning module is proposed. This module applies temporal perturbations and noise injection to the SSVEP signal to generate an enhanced view and utilizes a contrastive loss function to enhance the discriminative power and robustness of the model features.

[0050] Through the organic synergy of the above three technical modules, the problems of distribution shift and label dependency in cross-subject SSVEP classification are systematically solved, achieving high-precision classification without the need for labeled data of the target subjects.

[0051] More specifically, firstly, the formula for the proposal is defined as follows: Source domain with tags The source domain includes One EEG sample, Representing the One EEG sample ( This refers to the number of electrode channels. (number of sampling points) This represents the label corresponding to the sample. An unlabeled target domain is represented as... Due to significant individual differences among participants, the source and target domains follow different distributions. ).

[0052] Step 1: Filter Bank Euclidean Alignment: Considering the amplitude variability and inherent noise of EEG signals among different subjects, data alignment techniques are often used to reduce the marginal distribution shift between the source and target domains. Traditional Euclidean alignment typically operates at the channel level, achieving alignment by matching the covariance matrix of the EEG data. However, SSVEP signals contain non-negligible harmonic responses, requiring the capture of multi-band information through a filter bank. Channel-level Euclidean alignment may therefore overlook complementary information between different frequencies, such as... Figure 2 As shown. To overcome this limitation, a Euclidean alignment strategy for the filter bank is proposed. The signal after filter bank decomposition is denoted as... , Number of filter banks, reference matrix Defined as:

[0053] (1);

[0054] In the formula, This represents the average covariance matrix of all samples.

[0055] The alignment operation is defined as follows:

[0056] (2);

[0057] The average covariance matrix of the aligned signal is an identity matrix, achieving signal whitening. The proposed Euclidean alignment strategy for the filter bank utilizes the frequency information of the filter bank to achieve more accurate alignment and effectively reduce distribution offset.

[0058] Step 2, Cross-subject self-training module: To address the issues of mismatched source and target domain distributions and low pseudo-label quality, a two-stage cross-subject self-training module consisting of adversarial pre-training and dual ensemble self-training is designed.

[0059] Adversarial learning pre-training mechanism: Considering the spectral, spatial, and temporal characteristics of SSVEP signals, this invention employs a convolutional neural network (CNN)-based network G for feature extraction. This network integrates filter bank fusion, spatial filtering, and temporal feature extraction functions, and is paired with a classification head H. To align feature distributions across different domains, a domain discriminator D is introduced. The feature extraction network G and the domain discriminator D are connected through a gradient inversion layer, which inverts the gradient flowing through G. The loss function for the first stage, i.e., the pre-training stage, is defined as follows:

[0060] (3);

[0061] (4);

[0062] (5);

[0063] Dual-integration self-training mechanism: Traditional self-training methods are susceptible to low-quality pseudo-labels, and directly using these pseudo-labels leads to error accumulation. To address this issue, this invention proposes a dual-integration self-training strategy. This strategy first utilizes the mean teacher paradigm to iteratively optimize unlabeled target domain data. This paradigm employs two structurally identical models: a teacher model and a student model. The teacher model first predicts the target domain data and generates pseudo-labels through confidence level filtering. These pseudo-labels are then used to guide the student model's learning. The parameters of the teacher model... Through student model parameters The exponential moving average is used for updating, and its update formula is defined as:

[0064] (6);

[0065] In the formula, The momentum coefficient is the momentum coefficient of the exponential moving average. Therefore, the teacher model can be viewed as a temporal ensemble of the student model.

[0066] Furthermore, this invention proposes a pseudo-label fusion strategy to achieve the integration of multiple enhanced views. Let... Pseudo-labels representing the original data, and These represent the pseudo-labels obtained from the two enhanced views. The merged pseudo-label is defined as follows:

[0067] (7);

[0068] (8);

[0069] In the formula, Indicates projection features, Indicates projection relative to the original view Calculate the cosine similarity. To ensure numerical stability.

[0070] Based on the above-mentioned fused pseudo-labels During the self-training phase, the student model performs supervised optimization learning on samples from the target domain. Specifically, the student model uses the target domain input data and pseudo-labels generated by the teacher model as training supervision information, and updates its parameters by minimizing the classification loss function. Therefore, the classification loss for the target domain is defined as follows:

[0071] (9);

[0072] Step 3, Time-Frequency Enhanced Contrastive Learning: Although obtaining reliable pseudo-labels can significantly improve performance, self-training based on cross-entropy is still affected by noisy labels. Therefore, this invention proposes a time-frequency enhanced contrastive learning module to enhance feature discrimination capabilities. For steady-state visual evoked potential (SSVEP) EEG signals, two data augmentation methods are introduced: temporal perturbation and noise injection. Using these augmentation methods, the teacher model generates multiple predictions for SSVEP samples in the same target domain. These predictions are then fused using a weighted strategy to obtain more robust pseudo-labels. These augmentation operations are performed along the temporal and spectral dimensions to generate diverse views. Based on these views, a supervised contrastive loss is employed. This loss promotes the clustering of feature representations for similar samples while separating the feature representations for different samples. The contrastive loss is defined as:

[0073] (10);

[0074] In the formula, Representation and Sample Augmented sample sets with the same predicted class Represents the set of all augmented samples in the batch. This represents the temperature coefficient.

[0075] The loss function for the second stage is defined as follows:

[0076] (11);

[0077] This invention addresses the problems of large signal variability and high annotation costs in the SSVEP-BCI method across subjects. First, a filter bank Euclidean alignment strategy is designed, using filter bank-channel joint covariance matrix whitening to align signal frequency-spatial information. Then, a cross-subject self-training framework is constructed. In the pre-training phase, adversarial learning is used to align the source and target domain distributions. In the self-training phase, student-teacher models iterate, and time integration and multi-view pseudo-label fusion optimize pseudo-labels. A time-frequency data augmentation module is also introduced to enhance feature discriminative power. As shown in Tables 1-2, on the Benchmark dataset, this method achieves a maximum information transfer rate of 203.1 ± 8.03 bits / min at 0.8 seconds, significantly outperforming the current best method (194.54 ± 10.07 bits / min, p < 0.01). On the BETA dataset, it achieves a maximum information transfer rate of 160.93 ± 6.93 bits / min at 0.8 seconds, significantly higher than the SFDA method (131.99 ± 7.86 bits / min, p < 0.001).

[0078] Table 1: Accuracy and data transfer rate of this method on the Benchmark dataset

[0079] Time (s) 0.2 0.4 0.6 0.8 1.0 Accuracy (%) 24.36±1.97 57.9±3.15 76.54±2.89 89.51±2.44 94.8±1.5 ITR (bit / min) 49.20 ±6.73 148.08±11.44 185.6±9.94 203.1±8..03 193.12±4.86

[0080] Table 2: Accuracy and data transmission rate of this method on the BETA dataset

[0081] Time (s) 0.2 0.4 0.6 0.8 1.0 Accuracy (%) 23.37±1.11 46.84±2.42 64.97±2.46 77.19±2.28 82.91±1.94 ITR (bit / min) 44.72±3.52 109.30±8.39 145.94±8.0 160.93±6.93 155.32±5.41

[0082] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A self-training-based cross-subject domain adaptive SSVEP EEG classification method, characterized in that, include: Acquire the SSVEP signal, capture the multi-band information of the SSVEP signal, and perform Euclidean alignment of the filter bank. A convolutional neural network model was trained using the aligned SSVEP signal. The training process includes: a pre-training phase and a self-training phase; In the pre-training stage, the aligned SSVEP signal is input into a convolutional neural network model to obtain target features. The target features are classified using a classification head, and a domain discriminator is introduced to align the feature distributions of different domains to obtain a pre-trained EEG classification model. During the self-training phase, the pre-trained EEG classification model is divided into a teacher model and a student model. The teacher model is used to predict target domain data, and pseudo-labels are generated through confidence screening. Furthermore, a pseudo-label fusion strategy and a time-frequency enhanced contrastive learning strategy are introduced to guide the learning of the student model, thereby obtaining an EEG classification model. The EEG classification model is then used for SSVEP EEG classification.

2. The self-training-based cross-subject domain adaptive SSVEP EEG classification method according to claim 1, characterized in that, Capturing the multi-band information of the SSVEP signal and performing Euclidean alignment of the filter bank includes: ； in, For the aligned signal, The mean covariance matrix for all samples is... This is the original signal.

3. The self-training-based cross-subject domain adaptive SSVEP EEG classification method according to claim 1, characterized in that, The convolutional neural network model is connected to the domain discriminator through a gradient inversion layer, which can invert the gradients flowing through the convolutional neural network model.

4. The self-training-based cross-subject domain adaptive SSVEP EEG classification method according to claim 1, characterized in that, The loss function in the pre-training phase includes: ； ； ； in, For source domain classifier, The cross-entropy loss is the source domain. For classification header, For feature extractors, For the true label of the source domain, For the domain discriminator loss, For the trainable parameters of the feature extractor G, Let D be the trainable parameters of the domain discriminator. For the discrimination loss of the source domain, For source domain feature extractor, For the discrimination loss of the target domain, For domain discriminators.

5. The self-training-based cross-subject domain adaptive SSVEP EEG classification method according to claim 1, characterized in that, Obtaining the EEG classification model includes: The teacher model is used to predict target domain data, and pseudo-labels are generated by confidence screening. The pseudo-label fusion strategy is further introduced to integrate multiple augmented views to guide the learning of the student model and obtain the EEG classification model. While generating the pseudo-labels, the steady-state visual evoked potential signals under the pseudo-labels are subjected to temporal perturbation and noise injection along the time and spectral dimensions, respectively, to generate diversified views and enhance feature discrimination capabilities.

6. The self-training-based cross-subject domain adaptive SSVEP EEG classification method according to claim 5, characterized in that, Integrating the multiple enhanced views includes: ； ； in, Indicates projection features, Indicates projection relative to the original view Calculate the cosine similarity. The final prediction result after fusion of multiple enhanced views. To enhance the indexing of views, The weight coefficient for the k-th augmented view. This represents the model prediction result corresponding to the k-th augmented view. The projection feature vector obtained by passing the projector through the k-th augmented view. This is the projection feature vector corresponding to the original view. The projection feature vector obtained by passing the m-th enhanced view through the projector. This is a numerical stability term used to avoid the denominator being zero. This is the input sample for the k-th enhanced view.

7. The self-training-based cross-subject domain adaptive SSVEP EEG classification method according to claim 1, characterized in that, The learning process for the student model includes: The parameters of the teacher model are updated using an exponential moving average of the student model parameters: ； in, For the parameters of the teacher model, For student model parameters, The momentum coefficient of the exponential moving average.

8. The self-training-based cross-subject domain adaptive SSVEP EEG classification method according to claim 1, characterized in that, The loss function for the self-training phase includes: ； ； in, This is the loss function for the self-training phase. The loss of the student model for classifying samples in the target domain. The weight hyperparameters of the learning loss are used to adjust their proportion in the total loss. The contrastive learning loss is used with sample i as the anchor point. Let i be the set of positive samples. The sample index is the value in the set of positive samples. Let be the projected feature vector of sample i. This is the transpose of the vector. Let p be the projected feature vector of the positive sample p. For temperature coefficient, The sample index in the candidate sample set. Let i be the set of samples corresponding to anchor point sample i.

Images