Multi-view migration interpretable method based on soft variable embedding and discriminant structure preserving

Abstract

The invention discloses a multi-view migration interpretable method based on soft variable embedding and discriminant structure preserving, and particularly relates to the technical field of machine learning, and the method comprises the steps: obtaining a plurality of multi-view features for EEG samples of a tagged source domain and an untagged target domain through the extraction of a plurality of features; after TSK-FS antecedent network mapping, constructing inter-domain connection by adopting migration soft variable embedded consequent learning; a data structure is retained through a local-global structure retention item, so that a discriminant neighborhood relationship of original data is retained to the greatest extent in a migration process; and a target function is constructed in combination with a multi-view learning strategy, iterative optimization is performed through an enhanced Lagrange multiplier algorithm, and finally a result is output by using a given classifier.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and more specifically to a multi-view transfer interpretable method based on soft variable embedding and discriminative structure preservation. Background Technology

[0002] Motor imagery brain-computer interface (MI-BCI) systems decode intentions by performing pattern recognition on non-invasively acquired multi-channel EEG signals, showing broad application prospects in fields such as neurorehabilitation, prosthetic control, and human-computer interaction. However, traditional MI-EEG classification methods (such as those based on common spatial patterns (CSP) and LDA / SVM) only perform well with the same subject and the same dataset. They struggle to cope with the inherent non-stationarity of EEG signals, significant noise, and distributional differences between individuals and across sessions / datasets, leading to a significant decrease in model generalization performance and making it difficult to meet practical application requirements.

[0003] To address uncertainties and noise interference in EEG data, researchers have introduced fuzzy systems into the MI-EEG classification task. For example, existing techniques (T. Nguyen, I. Hettiarachchi, A. Khatami, et al. Classification of multi-class BCI data by common spatial pattern and fuzzy system[J]. IEEE Access, vol. 6, pp. 27873-27884, 2018.) utilize CSP combined with TSK fuzzy systems to mitigate signal interference. Existing techniques (E. Jiang, T. Huang, X. Yin. A combination of deep learning models and type-2 fuzzy for EEG motor imagery classification through spatiotemporal-frequency features[J]. Journal of Medical Engineering & Technology, vol.48, no.7, pp.262-275,2025.) propose a deep fuzzy architecture that combines Type-2 fuzzy activation functions with Bayesian hyperparameter optimization, significantly improving classification accuracy. However, most existing TSK fuzzy classifiers rely on single-domain data and directly regress the label space through least squares or gradient descent. The consequent parameter learning is prone to overfitting and lacks cross-domain adaptability.

[0004] Transfer learning, as an effective means to mitigate the distributional differences between subjects and sessions, has seen the emergence of various strategies: global domain alignment based on maximum mean difference (MMD), reducing marginal differences using Riemannian tangent space mapping, and dynamically constructing EEG channel functional connectivity graphs through graph convolutional networks. However, these methods typically only focus on global distribution or local geometric alignment, neglecting the preservation of class discriminative structures. They struggle to simultaneously consider both local neighborhood relationships and global discriminative information, thus limiting cross-domain performance improvements.

[0005] Meanwhile, multi-view learning significantly enhances MI-EEG classification performance by fusing various feature perspectives, including spatial domain (tangent space mapping), frequency domain (power spectral density), nonlinearity (multi-scale entropy), and deep learning (MLP). For example, cross-frequency interaction frameworks, multi-view sparse learning, and Euclidean / Riemannian manifold feature fusion methods effectively capture complementary information. However, these methods often rely on predefined frequency bands or channel subsets and static viewpoint weights, failing to dynamically balance the contributions of different perspectives in cross-domain scenarios, resulting in insufficient information utilization.

[0006] In summary, existing technologies each have their advantages in fuzzy classification, transfer learning, and multi-view learning, but they cannot simultaneously address the non-stationarity, uncertainty, and cross-domain generalization challenges of EEG signals within a unified framework. How to construct a cross-domain fuzzy classifier that can preserve discriminative structure, consider both local and global factors, and dynamically optimize multi-view contributions is precisely the technical problem this invention aims to solve. Summary of the Invention

[0007] To address this, the present invention provides a multi-view transfer interpretable method based on soft variable embedding and discriminative structure preservation. By closely integrating multi-view feature fusion with transfer learning and fuzzy reasoning, the present invention not only theoretically proposes a transfer soft variable embedding mechanism that takes into account both local geometric structure and global discriminative information, but also significantly improves the classification robustness and generalization ability of MI-EEG signals across subjects, sessions, and datasets in practice, thereby solving the problems raised in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a multi-view transfer interpretable method based on soft variable embedding and discriminative structure preservation, comprising the following steps: Step 1: For EEG samples with labeled source domain and unlabeled target domain, obtain multiple multi-view features through various feature extraction methods; Step 2: After mapping the TSK-FS preamble network, inter-domain connections are constructed using transfer soft variable embedded consequent learning; Step 3: Enhance intra-class compactness by modeling local neighborhood affinity between uniform samples in the fuzzy space using an intrinsic graph; the penalty graph represents the differences between dissimilar samples through exclusion constraints, thereby synergistically achieving local structure preservation and global discriminative enhancement; preserve data structure through discriminative structure to maximize the retention of discriminative neighborhood relationships of the original data during the transfer process; introduce a low-rank global regularization term to constrain the total rank of the source subclasses to achieve global data structure compactness in the soft variable space; integrate the local discriminative structure matrix and global low-rank constraints through local-global structure preservation terms to achieve discriminative feature learning. Step 4: By pairwise constraints on the parameters of the results from different perspectives, the coordinated optimization of the feature transformation coefficient matrix is ​​promoted. Shannon entropy is used to measure the weight of each perspective to achieve adaptive perspective fusion. The objective function is constructed by integrating consequent learning, local-global structure preservation and multi-perspective learning strategies. The result is then iteratively optimized by the enhanced Lagrange multiplier algorithm and finally output using the given classifier.

[0009] Preferably, step 1 specifically includes: The labeled source domain and the unlabeled target domain are each represented by M multi-view features using various feature extraction methods, denoted as follows: and The source domain and the m-th unlabeled view can be specifically represented as: and ,in and These represent the number of EEG samples in the two domains, respectively. This represents the sample set; the features of each viewpoint are mapped to a new feature space through the TSK-FS preflight network, according to the defined matrix. and Convert them into new features respectively and , where subscript and This corresponds to the tagged source domain and the untagged target domain.

[0010] Preferably, step 2 specifically includes: Traditional fuzzy consequent parameter learning methods directly map fuzzy features to a strict label space. To establish a connection between the source and target domains, and to enhance the flexibility of feature representation, consequent learning maps the source and target domains to a soft variable space. (1); in, It is a regularization parameter. Indicates the consequent parameter. Represents the label association matrix. The squared Frobenius norm is used to measure the difference between matrices; the first term in equation (12) guarantees the fidelity of the fuzzy feature mapping. It is a label matrix. It is the transpose of the label matrix; the second term supervises the soft variable matrix through the label matrix Y. Category separability; To bridge the differences between domains, a cross-domain linear representation is established: (2); in It is a representation matrix that connects the source domain and the target domain to establish inter-domain relationships; To capture more cross-domain relevance, The low-rank constraint is adopted; its purpose is: (1) to improve the compactness of cross-domain representation and achieve the de-redundancy of knowledge transfer by using only the most representative source samples to sparsely reconstruct the target samples; (2) to induce The block diagonal structure is formed, revealing the latent class structure in the soft variable space; in addition, since the number of classes is less than the number of training samples and the number of fuzzy rules, the soft variable matrix inherently satisfies the low-rank feature. The consequent learning representation constructed based on this is as follows: (3); in It is a regularization parameter. Let represent the rank of the matrix, 1 represent a vector of all 1s, and I represent the identity matrix.

[0011] Preferably, step 3 specifically includes: Preserving inherent local geometry and global discriminative information is crucial for improving cross-domain adaptability; therefore, this invention constructs... The inner diagram and punishment picture Inner Diagram Intra-class compactness is enhanced by modeling local neighborhood affinity between uniform samples in a fuzzy space, using a penalty graph. By using exclusion constraints to characterize the differences between dissimilar samples, local structure preservation and global discrimination enhancement can be achieved synergistically. Inner Diagram and punishment picture All are obtained based on the k-nearest neighbor criterion, and their elements and They are defined as follows: (4); (5); in, This represents the vector form of the j-th sample in the source domain s from the m-th viewpoint, where i and j are sample indices, using an intrinsic graph. and punishment picture This preserves the local geometry between cross-domain samples, ensuring that new features of similar samples remain close, while features of different samples are pushed apart; The local structure preservation term is represented as: (6); in, Represents the trace of a matrix; It is the transpose of the consequent parameter; preserving the global structure of the data is another key goal of subsequent parameter learning; in order to achieve a compact global data structure in the soft variable space, a low-rank global regularization term is introduced to constrain the total rank of the source subclass to approximate the global data rank in the soft variable space: (7); First, by employing global low-rank consistency, it inherently preserves the low-dimensional manifold structure, which is particularly important for processing unlabeled target domain data; second, by enforcing subclass-specific low-rank reconstruction within the source domain, it effectively reduces intra-class distance and promotes separation between subclasses. Therefore, the local-global structure preservation term is: (8); in, It represents the fuzzy feature matrix that maps the sample set of the k-th subclass under the m-th view in the source domain s to the new feature space; This represents the set of features after mapping the source and target domains; definition , This represents the local discriminant structure matrix of the m-th viewpoint (m theoretically must be greater than or equal to 1, but in practice, it is greater than or equal to 2; in this invention, m is at most 4), and is the intrinsic graph Laplacian matrix. With the penalty graph Laplace matrix The difference embodies the constraints of local intra-class compactness and inter-class separation; Equation (8) is: (9).

[0012] Preferably, step 4 specifically includes: Using pairwise constraints on the result parameters from different perspectives to facilitate the eigentransform coefficient matrix Coordination optimization: (10); Then, a viewpoint weight penalty term based on Shannon entropy is used to measure the weight of each viewpoint: (11); in It is the viewpoint weight matrix; Based on consequent learning through transfer soft variable embedding, local-global structure preservation, and multi-view learning strategies, the objective function of the multi-view transfer TSK fuzzy classifier MVT-TSK-SVSP is expressed as: (12); in , and It is a regularization parameter; Due to the nonconvexity and discreteness of the rank function, it is difficult to directly optimize equation (12). Therefore, the nuclear norm is used to replace the rank operation; equation (12) is written as: (13); According to the definition of nuclear norm, ,in , . , in , , in . To facilitate model solving, auxiliary variables are added. and Equation (13) can be written as: (14); Equation (14) is solved using the enhanced Lagrange multiplier algorithm; the unconstrained optimization problem of equation (14) is written as follows: (15); The optimization process using the iterative strategy is as follows; 1) Update The optimization of equation (15) can be simplified into the following subproblems: (16); get: (17); 2) Update The optimization of equation (15) can be simplified into the following subproblems: (18); After rearrangement, equation (18) can be expressed as: (19); get: (20); Considering constraints Due to column normalization constraints , Represented as . 3) Update The optimization of equation (15) can be simplified into the following subproblems: (twenty one); It can be obtained, (twenty two); 4) Update The optimization of equation (15) is simplified into the following subproblem: Update : (twenty three); in It is a penalty parameter; get: (twenty four); 5) Update The optimization of equation (15) can be simplified into the following subproblems: (25); Let

[0013] get: (26); 6) Update the Lagrange multipliers: (27); After iterating through parameters 1)-6) of the iterative strategy, the consequent parameters are obtained. and perspective weight The optimal solution for unlabeled target domain samples; The final output result is The decision function is expressed as follows: ; As a given classifier, this invention uses the nearest neighbor classifier.

[0014] This invention dynamically weights multi-channel spatial features, frequency domain power spectrum, nonlinear entropy features, and complementary information extracted from deep learning perspectives, overcoming the shortcomings of traditional single-viewpoint methods in modeling the interaction of signal diversity and complexity. Secondly, through low-rank graph embedding and structure-preserving constraints, this invention maximizes the preservation of the original data's discriminative neighborhood relationships during the transfer process, avoiding the degradation risk caused by conventional domain alignment methods neglecting class structure. Thirdly, the introduced soft variable embedding enables a smooth transition and adaptive correction of source domain knowledge in the target domain, significantly reducing reliance on large amounts of labeled data, shortening system calibration time, and reducing user fatigue. Furthermore, the defuzzification output of the TSK fuzzy system combines nonlinear mapping with highly interpretable rule reasoning, facilitating real-time decoding and feedback control, providing efficient, stable, and transparent decision support for closed-loop BCI applications. Compared to existing technologies, this invention not only technically overcomes the challenges of data distribution differences and uncertainties faced by existing MI-BCI cross-domain classification but also lays a solid foundation for the subsequent development of intelligent neurorehabilitation, prosthetic control, and human-machine collaborative systems based on fuzzy transfer and multi-viewpoint learning, possessing significant theoretical value and broad application prospects.

[0015] Specifically, it includes: 1. Multi-view feature soft variable embedding mechanism For the first time, spatial domain (CSP), frequency domain (power spectral density), nonlinear entropy features, and depth features are uniformly mapped to the soft variable space of the Takagi–Sugeno–Kang fuzzy system. By dynamically adjusting the contribution of each perspective in classification decision through adaptive weight learning, the utilization efficiency of complementary information of MI-EEG signals is greatly improved.

[0016] 2. Local-Global Structure Preservation in Low-Rank Graph Embedding We propose a multi-view local neighborhood and global discriminative structure preservation method based on low-rank graph regularization. This method not only aligns the marginal distribution between the source and target domains but also strictly maintains intra-class similarity and inter-class separability, enabling seamless transfer of cross-domain discriminative structures.

[0017] 3. Perspective Consistency Cross-Domain Constraints The innovative introduction of a perspective consistency loss term constrains the consistency of soft variable representations from different perspectives on the same sample, effectively suppressing overfitting from a single perspective and significantly improving model robustness in cross-subject / cross-session scenarios.

[0018] 4. Reuse of consequent parameters driven by migration soft variables In TSK consequent parameter learning, the parameter training complexity in the target domain with small sample size is reduced by similarity-weighted transfer of the consequent parameters already learned in the source domain, and the consequent output is further corrected by soft variable embedding, thus breaking through the bottleneck of traditional least squares regression being prone to overfitting.

[0019] 5. End-to-end alternating iterative optimization strategy This invention designs an efficient iterative algorithm that combines alternating minimization and closed-form update of the solution. It can simultaneously optimize the fuzzy antecedent (membership parameter) and consequent (linear regression parameter), avoiding the common problems of slow convergence and easy getting trapped in local optima in gradient descent, and achieving both fast convergence and high accuracy.

[0020] 6. Real-time online decoding and incremental update capabilities For closed-loop BCI application scenarios, a method for rapid updating of perspective weights and consequent parameters based on incremental learning is proposed to enable online fine-tuning when new subjects or new session data arrive, ensuring the system's ability to respond instantly to changes in the environment and user status.

[0021] 7. Wide adaptability and scalability This invention features a modular design that allows for seamless integration with other feature extractors (such as figure neural networks and time-frequency joint convolutional networks) and domain adaptation techniques. It is easily extended to scenarios such as multi-task or multi-class MI-EEG classification and cross-domain decoding of other bioelectric signals (such as MEG and sEMG).

[0022] Together, these technologies construct an end-to-end MI-EEG classification framework that balances multi-perspective information fusion, discriminative structure preservation, and efficient cross-domain adaptation. This framework significantly improves the generalization performance and real-time performance of brain-computer interface systems in real-world environments, meeting the technical requirements for high-performance, low-latency, and cross-domain robust decoding in motor imagery tasks. Attached Figure Description

[0023] Figure 1 This is a schematic diagram illustrating the fuzzy rule comparison of the BCI-2a dataset provided by the present invention. Figure 2 This is a schematic diagram illustrating the fuzzy rule comparison of the OpenBMI dataset provided by the present invention; Figure 3 This is a schematic diagram illustrating the comparison of fuzzy rules between BCI-2a and OpenBMI provided by the present invention. Figure 4 This is a schematic diagram illustrating the comparison of fuzzy rules between OpenBMI and BCI-2a provided by the present invention. Detailed Implementation

[0024] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0025] The TSK fuzzy model consists of three parts: antecedent, consequent, and rule base. TSK can handle fuzzy inputs and fuzzy outputs and make decisions through fuzzy reasoning. In classification problems, a multi-input... Multiple outputs The rule is expressed as, (1); Where r (r=1,…,R) represents the number of fuzzy rules. K and d represent the number of categories and dimension of the sample, respectively.

[0026] The antecedent of a rule is a fuzzy set of input variables. The fuzzy set of the r-th rule in the j-th dimension is represented as follows: The consequent of a rule is represented by the deterministic value of the output variable, and the consequent parameter... , It is the subsequent parameter vector of the k-th output in the r-th rule.

[0027] In the first layer of the model, the TSK model calculates the membership degree of each fuzzy set in the antecedent based on the input variables. Using a Gaussian function, the membership function of the m-th fuzzy rule in the j-th dimension is... Represented as: (2); in and These are the center and width of the membership function, respectively.

[0028] The second layer of the model is responsible for calculating the trigger strength of the fuzzy rules. : (3); The third layer of the model is responsible for calculating the trigger strength of the defuzzified fuzzy rules. When using the normalization method, we can obtain: (4); The fourth layer of the model is responsible for calculating the output (fuzzy features) of the fuzzy rules. Based on the trigger strength and consequent parameters of the fuzzy rules, we can obtain... (5); Define the following matrix (6); (7); (8); The output of the 5th layer of the model is represented as follows: (9); For classification problems, the class label of the test sample z is represented by the K output nodes. The category corresponding to the component with the largest median value.

[0029] The TSK fuzzy model requires parameter optimization involving antecedent parameters. And the consequent parameter C. The antecedent parameter is often obtained using clustering methods. If fuzzy C-means clustering (FCM) is used, Equal to FCM cluster center, The formula for calculation is, (10); in The fuzzy membership degree is obtained from clustering.

[0030] Traditional TS fuzzy models often use gradient descent to optimize antecedent and consequent parameters. In recent years, some scholars have used clustering to optimize antecedent parameters and least squares to optimize consequent parameters. This method has the characteristics of strong versatility, good generalization ability and strong stability.

[0031] From equation (4), the mapping of the original dataset X in the fuzzy space is: ,in Then, based on the least squares loss function, the optimized representation of the consequent parameters of the TS fuzzy model is as follows: (11); in It is the regularization parameter.

[0032] This invention provides a multi-view transfer interpretable method based on soft variable embedding and discriminative structure preservation, comprising the following steps: Step 1: For EEG samples with labeled source domain and unlabeled target domain, obtain multiple multi-view features through various feature extraction methods; In the MVT-TSK-SVSP model, the labeled source domain and the unlabeled target domain are represented by M multi-view features through various feature extraction methods, denoted as follows: and The source domain and the m-th unlabeled view can be specifically represented as: and ,in and These represent the number of EEG samples in the two domains, respectively. The sample set is represented; the features of each viewpoint are mapped to a new feature space through the TSK-FS preflight network, according to equation (8). and Convert them into new features respectively and , where subscript and This corresponds to the tagged source domain and the untagged target domain.

[0033] Step 2: After mapping the TSK-FS preamble network, inter-domain connections are constructed using transfer soft variable embedded consequent learning; Traditional fuzzy consequent parameter learning methods directly map fuzzy features to a strict label space. To establish a connection between the source and target domains, and to enhance the flexibility of feature representation, consequent learning maps the source and target domains to a soft variable space. (12); in, It is a regularization parameter. Indicates the consequent parameter. Represents the label association matrix. The squared Frobenius norm is used to measure the difference between matrices; the first term in equation (12) guarantees the fidelity of the fuzzy feature mapping. It is a label matrix. It is the transpose of the label matrix; the second term supervises the soft variable matrix through the label matrix Y. Category separability; To bridge the differences between domains, a cross-domain linear representation is established: (13); in It is a representation matrix that connects the source domain and the target domain to establish inter-domain relationships; To capture more cross-domain relevance, The low-rank constraint is adopted; its purpose is: (1) to improve the compactness of cross-domain representation and achieve the de-redundancy of knowledge transfer by using only the most representative source samples to sparsely reconstruct the target samples; (2) to induce The block diagonal structure is formed, revealing the latent class structure in the soft variable space; in addition, since the number of classes is less than the number of training samples and the number of fuzzy rules, the soft variable matrix inherently satisfies the low-rank feature. The consequent learning representation constructed based on this is as follows: (14); in It is a regularization parameter. Let represent the rank of the matrix, 1 represent a vector of all 1s, and I represent the identity matrix.

[0034] Step 3: Enhance intra-class compactness by modeling local neighborhood affinity between uniform samples in the fuzzy space using an intrinsic graph; the penalty graph represents the differences between dissimilar samples through exclusion constraints, thereby synergistically achieving local structure preservation and global discriminative enhancement; preserve data structure through discriminative structure to maximize the retention of discriminative neighborhood relationships of the original data during the transfer process; introduce a low-rank global regularization term to constrain the total rank of the source subclasses to achieve global data structure compactness in the soft variable space; integrate the local discriminative structure matrix and global low-rank constraints through local-global structure preservation terms to achieve discriminative feature learning. Preserving inherent local geometry and global discriminative information is crucial for improving cross-domain adaptability; therefore, this invention constructs... The inner diagram and punishment picture Inner Diagram Intra-class compactness is enhanced by modeling local neighborhood affinity between uniform samples in a fuzzy space, using a penalty graph. By using exclusion constraints to characterize the differences between dissimilar samples, local structure preservation and global discrimination enhancement can be achieved synergistically. Inner Diagram and punishment picture All are obtained based on the k-nearest neighbor criterion, and their elements and They are defined as follows: (15); (16); in, This represents the vector form of the j-th sample in the source domain s from the m-th viewpoint, where i and j are sample indices, using an intrinsic graph. and punishment picture This preserves the local geometry between cross-domain samples, ensuring that new features of similar samples remain close, while features of different samples are pushed apart; The local structure preservation term is represented as: (17); in, Represents the trace of a matrix; It is the transpose of the consequent parameter; Preserving the global structure of the data is another key objective for subsequent parameter learning. To achieve a compact global data structure in the soft variable space, a low-rank global regularization term is introduced to constrain the total rank of the source subclasses to approximate the global data rank in the soft variable space. (18); First, by employing global low-rank consistency, it inherently preserves the low-dimensional manifold structure, which is particularly important for processing unlabeled target domain data; second, by enforcing subclass-specific low-rank reconstruction within the source domain, it effectively reduces intra-class distance and promotes separation between subclasses. Therefore, the local-global structure preservation term is: (19); in, It represents the fuzzy feature matrix that maps the sample set of the k-th subclass under the m-th view in the source domain s to the new feature space; This represents the set of features after mapping the source and target domains; definition , This represents the local discriminant structure matrix of the m-th viewpoint, which is the intrinsic graph Laplacian matrix. With the penalty graph Laplace matrix The difference embodies the constraints of local intra-class compactness and inter-class separation; Equation (19) is: (20).

[0035] Step 4: By pairwise constraints on the parameters of the results from different perspectives, the coordinated optimization of the feature transformation coefficient matrix is ​​promoted. Shannon entropy is used to measure the weight of each perspective to achieve adaptive perspective fusion. The objective function is constructed by integrating consequent learning, local-global structure preservation and multi-perspective learning strategies. The result is then iteratively optimized by the enhanced Lagrange multiplier algorithm and finally output using the given classifier.

[0036] To effectively utilize the different features of MI-EEG and mitigate potential information loss in transfer learning, the designed fuzzy classifier introduces a multi-view learning mechanism, which enhances the semantic consistency between views and dynamically balances the view-specific contributions.

[0037] In particular, pairwise constraints on the result parameters from different perspectives are used to facilitate the transformation coefficient matrix. Coordination optimization: (twenty one); Then, a viewpoint weight penalty term based on Shannon entropy is used to measure the weight of each viewpoint: (twenty two); in It is the viewpoint weight matrix; Based on consequent learning through transfer soft variable embedding, local-global structure preservation, and multi-view learning strategies, the objective function of the multi-view transfer TSK fuzzy classifier MVT-TSK-SVSP is expressed as: (twenty three); in , and It is a regularization parameter; Due to the nonconvexity and discreteness of the rank function, it is difficult to directly optimize equation (23). Therefore, the nuclear norm is used to replace the rank operation; equation (23) is written as: (twenty four); According to the definition of nuclear norm, ,in , . , in , , in . To facilitate model solving, auxiliary variables are added. and Equation (24) can be written as: (25); Equation (25) is solved using the enhanced Lagrange multiplier algorithm; the unconstrained optimization problem of equation (25) is written as follows: (26); The optimization process using the iterative strategy is as follows; 1) Update The optimization of equation (26) can be simplified into the following subproblems: (27); get: (28); 2) Update The optimization of equation (26) can be simplified into the following subproblems: (29); After rearrangement, equation (29) can be expressed as: (30); get: (31); Considering constraints Due to column normalization constraints , Represented as . 3) Update The optimization of equation (26) can be simplified into the following subproblems: (32); It can be obtained, (33); 4) Update The optimization of equation (26) is simplified into the following subproblem: Update : (34); in It is a penalty parameter; get: (35); 5) Update The optimization of equation (26) can be simplified into the following subproblems: (36); Let

[0038] get: (37); 6) Update the Lagrange multipliers: (38); After iterating through parameters 1)-6) of the iterative strategy, the consequent parameters are obtained. and perspective weight The optimal solution for unlabeled target domain samples; The final output result is The decision function is expressed as follows: ; As a given classifier, this invention uses the nearest neighbor classifier.

[0039] This embodiment aims to comprehensively and objectively evaluate the classification performance of the MVT-TSK-SVSP model of the present invention under different human and dataset conditions and configurations in MI-EEG. The following evaluation metrics were mainly used: (1) Classification Accuracy. This embodiment uses accuracy (ACC) as the performance metric for the model. Average accuracy values ​​are given for all cross-human (BCI-2a, OpenBMI) and cross-dataset (BCI-2a→OpenBMI, OpenBMI→BCI-2a) experiments to measure the overall performance of different methods in various scenarios. Accuracy refers to the proportion of correctly predicted samples out of the total number of samples, reflecting the overall classification performance of the model. Its calculation formula is: ; Wherein, TP (True Positive) represents the number of samples correctly predicted as positive by the model; TN (True Negative) represents the number of samples correctly predicted as negative by the model; FP (False Positive) represents the number of samples incorrectly predicted as positive by the model; and FN (False Negative) represents the number of samples incorrectly predicted as negative by the model.

[0040] (2) Cohen's Kappa coefficient (Kappa). A statistical metric that measures the consistency between the model's classification results and random classifications. It can eliminate the bias caused by accidental correctness, thus reflecting the reliability of the classifier more objectively. Kappa values ​​were reported in cross-human and cross-dataset experiments to supplement accuracy evaluation.

[0041] The experimental environment is a fundamental condition for conducting experiments. This embodiment describes the experimental environment in detail as follows:

[0042] The comparison method in this embodiment is divided into three groups: multi-view learning, transfer learning, and multi-view transfer learning, including: 1) MVFFR: A multi-view EEG representation method that combines multiple features with a multi-view feature fusion strategy.

[0043] 2) LR-CR²VS-TSK: A multi-view TSK fuzzy system that optimizes fuzzy rule generation and subsequent parameter selection by utilizing label relaxation and dual sparse regularization (cross rules and cross perspectives).

[0044] 3) METL: Transfer learning based on manifold embedding, which synchronizes domain geometry alignment with classification parameter optimization.

[0045] 4) FWR-JPDA: A transfer learning method that uses a dual regularization strategy to integrate domain adaptation into feature extraction. It addresses distribution differences in cross-domain MI-EEG scenarios.

[0046] 5) MMDA: Multi-manifold alignment transfer learning enhances cross-disciplinary classification by maximizing intra-class compactness and inter-class separability.

[0047] 6) MVTL-LSR: Multi-view transfer learning with latent space regularization, which ensures multi-view complementarity and cross-human knowledge transfer through consistency constraints.

[0048] 7) O-MV-T-TSK-FS: Online multi-view transmission TSK fuzzy system dynamically optimizes fuzzy rules and aligns cross-domain feature distributions to reduce individual differences.

[0049] In the experiment, Search for regularization parameters within the range. The number of fuzzy rules in each viewpoint is searched within the range. The weighting factor is set to 2. The corresponding parameters in the comparison algorithm use their default settings. This embodiment uses classification accuracy and Kappa coefficient to evaluate model performance.

[0050] A. Cross-person comparison experiment This embodiment conducts a cross-subject comparison experiment by designating the MI-EEG data of one subject as the target domain and the data of all remaining subjects as the source domain. Experimental results on the BCI-2a and OpenBMI datasets are summarized in Tables I and II, respectively. Evaluation on both datasets shows that the proposed TSK classifier performs excellently in terms of both mean accuracy and kappa coefficient. Specifically, on the BCI-2a dataset, the classifier in this embodiment improves accuracy by 11.27% and kappa gain by 0.1296 compared to the MVFFR method, while also improving accuracy and kappa by 2.17% and 0.0320 respectively compared to the O-MV-T-TSK-FS method. Similar results were observed on the OpenBMI dataset, where the classifier in this embodiment achieves 7.63% higher accuracy and 0.1108 higher kappa than MVFFR, and has an advantage of 2.26% higher accuracy and 0.0175 kappa over O-MV-T-TSK-FS.

[0051] Comparative analysis shows that multi-view transfer learning methods consistently outperform traditional multi-view learning or traditional transfer learning, which are single-modal methods. Traditional multi-view learning suffers from training bias due to differences in domain distribution and limited target domain samples, while traditional transfer learning struggles to handle linear feature concatenation, neglecting nonlinear interactions and exacerbating dimensionality challenges. In contrast, the proposed TSK classifier integrates fuzzy classification, cross-domain adaptation, and multi-view learning. First, it replaces the traditional maximum mean discrepancy technique with inter-domain linear representations by integrating graph embeddings and low-rank constraints. This joint modeling framework simultaneously captures local geometric relationships and global structural consistency in the data. Second, it incorporates an adaptive weighting mechanism that leverages the unique discriminative power of multiple feature perspectives while ensuring consistency and collaborative advantages between perspectives.

[0052] Table I. Classification accuracy (%) and kappa for cross-human experiments on the BCI-2a dataset

[0053] Table II. Classification accuracy (%) and kappa for cross-human experiments on the OpenBMI dataset.

[0054] B. Cross-dataset experiments This embodiment conducts a cross-dataset comparison experiment by designating the EEG data of all subjects in one dataset (such as BCI-2a or OpenBMI) as the source domain and the subject data in another dataset as the target domain. Twenty-one subjects were selected for evaluation. To achieve cross-dataset classifier training, systematic alignment processing was applied to the original MI-EEG data, including class space alignment (retaining only left-handed and right-handed data in the BCI-2a dataset to ensure consistency in binary classification) and spatial topology coordination (selecting 22 common electrode channels based on compatibility analysis of EEG cap configuration). Tables III and IV summarize the experimental results of BCI-2a versus OpenBMI and OpenBMI versus BCI-2a, respectively. As shown in Tables III and IV, compared to the cross-subject results in Tables I and II, the classification accuracy and Kappa of all algorithms decreased. This performance degradation is due to significant cross-dataset differences in EEG acquisition equipment, experimental paradigms, subject physiology, and environmental conditions, coupled with inherent inter-subject variability, which amplifies the challenge of EEG signal recognition. Despite these complexities, the proposed TSK fuzzy classifier achieves the highest classification accuracy and Kappa. Compared to the lowest performing MVFFR, our classifier improves the average accuracy by 11.13% and the Kappa by 0.1296, while also outperforming the second-best performing O-MV-T-TSK-FS by 1.78% and 0.0280 in accuracy and Kappa, respectively.

[0055] Table III. Classification accuracy (%) and kappa in cross-dataset experiments from BCI-2a to OpenBMI

[0056] Table IV. Classification accuracy (%) and kappa in the OpenBMI to BCI-2a cross-dataset experiment.

[0057] C. Fuzzy rule quantity analysis The number of rules largely determines the performance of a TSK fuzzy classifier, as it directly shapes the model's interpretability and generalization ability. Since O-MV-T-TSK-FS is a multi-view transfer TSK model, this embodiment compares it with the classifier in this embodiment. Their classification accuracy is compared with the number of rules, which is adjusted incrementally from 2 to 12. Experimental results are as follows... Figures 1-4As shown in the example, the classification accuracy varies with the number of rules. The MVT-TSK-SVSP classifier in this example achieves optimal accuracy with fewer rules, demonstrating its ability to balance simplicity and effectiveness.

[0058] The proposed MVT-TSK-SVSP classifier offers a breakthrough and effective approach to solving the MI-EEG classification challenge. By cleverly combining fuzzy classifiers, transfer learning, and multi-view learning techniques, this classifier achieves cross-domain distribution alignment, ensures multi-view consistency, and leverages complementary information to collectively improve the performance of MI-EEG classification. At the heart of the framework is the result learning mechanism of transfer soft variable embedding, which replaces traditional TSK rule-based label regression with discriminative feature space construction. The proposed method significantly reduces the risk of overfitting in cross-domain scenarios while preserving task-relevant information. Furthermore, the low-rank constraint local-global structure preservation mechanism of the embedding graph captures domain-invariant global relevance and fine-grained local geometric patterns, ensuring robust knowledge transfer. The multi-view learning component combines maximum entropy regularization and view consistency constraints to dynamically optimize view-specific contributions while maintaining inter-view consistency.

[0059] This embodiment utilizes two MI-EEG datasets: BCI Competitive IV-2a (BCI-2a) and OpenBMI. The BCI-2a dataset contains MI-EEG data from nine healthy subjects who performed four different motor imagery tasks (left hand, right hand, foot, and tongue movements). EEG signals were recorded using 22-channel Ag / AgCl electrode caps at a sampling rate of 250 Hz. Each subject attended two sessions on different days, each session consisting of six runs. Data were extracted from the 2–6 second intervals of the MI execution phase and downsampled to 250 Hz for subsequent analysis.

[0060] The OpenBMI dataset comprises MI-EEG data from 54 participants performing two types of MI tasks (left-hand and right-hand). Each participant completed two phases, each involving 400 trials (100 trials per hand). A single trial consisted of a 3-second preparation phase and a 4-second execution phase. EEG signals were recorded at a raw sampling rate of 1000 Hz using 62-channel electrode caps and then downsampled to 250 Hz. Similar to BCI-2a, the OpenBMI dataset analyzed data at 4-second intervals, covering both the preparation and execution phases.

[0061] This experiment employs a multidimensional feature extraction method to comprehensively characterize EEG signals from four complementary perspectives. For spatial domain features, tangent space mapping (TSM) is used to analyze the spatial relationships between electrode channels, capturing geometric patterns in the covariance structure of the MI signal. For frequency domain features, power spectral density (PSD) quantifies the energy distribution across physiologically relevant frequency bands, revealing the oscillatory characteristics of the MI process. For nonlinear features, nonlinear dynamics are evaluated using multiscale entropy (MDE), measuring changes in signal complexity across time scales to detect subtle neural state transitions. For deep learning functionality, a multilayer perceptron (MLP) learns hierarchical temporal patterns from the raw EEG signal.

[0062] This invention, by closely integrating multi-view feature fusion with transfer learning and fuzzy reasoning, not only theoretically proposes a transfer soft variable embedding mechanism that balances local geometric structure and global discriminative information, but also significantly improves the classification robustness and generalization ability of MI-EEG signals across subjects, sessions, and datasets in practice. First, multi-channel spatial features, frequency domain power spectrum, nonlinear entropy features, and complementary information extracted from deep learning perspectives are dynamically weighted, addressing the shortcomings of traditional single-view methods in modeling the interaction of signal diversity and complexity. Second, through low-rank graph embedding and structure-preserving constraints, this invention can maximally preserve the discriminative neighborhood relationships of the original data during the transfer process, avoiding the degradation risk caused by conventional domain alignment methods ignoring class structure. Third, the introduced soft variable embedding enables a smooth transition and adaptive correction of source domain knowledge in the target domain, greatly reducing dependence on large amounts of labeled data, shortening system calibration time, and reducing user fatigue. Furthermore, the defuzzification output of the TSK fuzzy system combines nonlinear mapping with highly interpretable rule reasoning, facilitating real-time decoding and real-time feedback control, providing efficient, stable, and transparent decision support for closed-loop BCI applications. In summary, this invention not only overcomes the challenges of data distribution differences and uncertainties faced by existing MI-BCI cross-domain classification, but also lays a solid foundation for the subsequent development of intelligent neurorehabilitation, prosthetic control, and human-machine collaborative systems based on fuzzy transfer and multi-view learning. It has significant theoretical value and broad application prospects.

[0063] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A multi-view transfer interpretable method based on soft variable embedding and discriminative structure preservation, characterized by: Includes the following steps: Step 1: For EEG samples with labeled source domain and unlabeled target domain, obtain multiple multi-view features through various feature extraction methods; Step 2: After mapping the TSK-FS preamble network, inter-domain connections are constructed using transfer soft variable embedded consequent learning; Step 3: Enhance intra-class compactness by modeling the local neighborhood affinity between uniform samples in the fuzzy space through the intrinsic graph, and characterize the differences between dissimilar samples through the exclusion constraint of the penalty graph, thereby synergistically achieving local structure preservation and global discrimination enhancement; By preserving the data structure through discriminative structure, the discriminative neighborhood relationships of the original data are preserved to the greatest extent during the migration process; a low-rank global regularization term is introduced to constrain the total rank of the source subclasses to achieve the compactness of the global data structure in the soft variable space; and a local-global structure preservation term is used to integrate the local discriminative structure matrix and the global low-rank constraint to achieve discriminative feature learning. Step 4: By pairwise constraints on the parameters of the results from different perspectives, the coordinated optimization of the feature transformation coefficient matrix is ​​promoted. Shannon entropy is used to measure the weight of each perspective to achieve adaptive perspective fusion. The objective function is constructed by integrating consequent learning, local-global structure preservation and multi-perspective learning strategies. The result is then iteratively optimized by the enhanced Lagrange multiplier algorithm and finally output using the given classifier.

2. The multi-view transfer interpretable method based on soft variable embedding and discriminative structure preservation according to claim 1, characterized in that: Step 1 specifically includes: The labeled source domain and the unlabeled target domain are each represented by M multi-view features using various feature extraction methods, denoted as follows: and The source domain and the m-th unlabeled view can be specifically represented as: and ,in and These represent the number of EEG samples in the two domains, respectively. This represents the sample set; the features of each viewpoint are mapped to a new feature space through the TSK-FS preflight network, according to the defined matrix. and Convert them into new features respectively and , where subscript and This corresponds to the tagged source domain and the untagged target domain.

3. The multi-view transfer interpretable method based on soft variable embedding and discriminative structure preservation according to claim 2, characterized in that: Step 2 specifically includes: To establish a connection between the source and target domains, and to enhance the flexibility of feature representation, consequent learning maps the source and target domains to a soft variable space: (1)； in, It is a regularization parameter. Indicates the consequent parameter. Represents the label association matrix. The square of the Frobenius norm is used to measure the difference between matrices; the first term in equation (1) guarantees the fidelity of the fuzzy feature mapping. It is a label matrix. It is the transpose of the label matrix; the second term supervises the soft variable matrix through the label matrix Y. Category separability; To bridge the differences between domains, a cross-domain linear representation is established: (2)； in It is a representation matrix that connects the source domain and the target domain to establish inter-domain relationships; The consequent learning representation constructed based on this is as follows: (3)； in It is a regularization parameter. Let represent the rank of the matrix, 1 represent a vector of all 1s, and I represent the identity matrix.

4. The multi-view transfer interpretable method based on soft variable embedding and discriminative structure preservation according to claim 3, characterized in that: Step 3 specifically includes: Inner Diagram and punishment picture All are obtained based on the k-nearest neighbor criterion, and their elements and They are defined as follows: (4)； (5)； in, This represents the vector form of the j-th sample in the source domain s from the m-th viewpoint, where i and j are sample indices, using an intrinsic graph. and punishment picture This preserves the local geometry between cross-domain samples, ensuring that new features of similar samples remain close, while features of different samples are pushed apart; The local structure preservation term is represented as: (6)； in, Represents the trace of a matrix; It is the transpose of the consequent parameter; to achieve a compact global data structure in the soft variable space, a low-rank global regularization term is introduced to constrain the total rank of the source subclass to approximate the global data rank in the soft variable space: (7)； Therefore, the discriminant structure, i.e., the local-global structure preservation term, is: (8)； in, It represents the fuzzy feature matrix that maps the sample set of the k-th subclass under the m-th view in the source domain s to the new feature space; This represents the set of features after mapping the source and target domains; definition , This represents the local discriminant structure matrix of the m-th viewpoint, where m is greater than or equal to 1, and is the intrinsic graph Laplacian matrix. With the penalty graph Laplace matrix The difference embodies the constraints of local intra-class compactness and inter-class separation; Equation (8) is: (9)。 5. The multi-view transfer interpretable method based on soft variable embedding and discriminative structure preservation according to claim 4, characterized in that: Step 4 specifically includes: Using pairwise constraints on the result parameters from different perspectives to facilitate the eigentransform coefficient matrix Coordination optimization: (10)； Then, a viewpoint weight penalty term based on Shannon entropy is used to measure the weight of each viewpoint: (11)； in It is the viewpoint weight matrix; Based on consequent learning through transfer soft variable embedding, local-global structure preservation, and multi-view learning strategies, the objective function of the multi-view transfer TSK fuzzy classifier MVT-TSK-SVSP is expressed as: (12)； in , and It is a regularization parameter; Due to the nonconvexity and discreteness of the rank function, it is difficult to directly optimize equation (12). Therefore, the nuclear norm is used to replace the rank operation; equation (12) is written as: (13)； According to the definition of nuclear norm, ,in , . , in , , in . To facilitate model solving, auxiliary variables are added. and Equation (13) can be written as: (14)； Equation (14) is solved using the enhanced Lagrange multiplier algorithm; the unconstrained optimization problem of equation (14) is written as follows: (15)； in, , , All are Lagrange multipliers; The optimization process using the iterative strategy is as follows; 1) Update The optimization of equation (15) can be simplified into the following subproblems: (16)； get: (17)； 2) Update The optimization of equation (15) can be simplified into the following subproblems: (18)； After rearrangement, equation (18) can be expressed as: (19)； get: (20)； Considering constraints Due to column normalization constraints , Represented as . 3) Update The optimization of equation (15) can be simplified into the following subproblems: (21)； It can be obtained, (22)； 4) Update The optimization of equation (15) is simplified into the following subproblem: Update : (23)； in It is a penalty parameter; get: (24)； 5) Update The optimization of equation (15) can be simplified into the following subproblems: (25)； Let ; get: (26)； 6) Update the Lagrange multiplier: (27)； After iterating through parameters 1)-6) of the iterative strategy, the consequent parameters are obtained. and perspective weight The optimal solution for unlabeled target domain samples; The final output result is The decision function is expressed as follows: ; As a given classifier.

Images