Multi-Riemann kernel fusion feature-based motor imagery electroencephalogram signal classification method, system and equipment and medium

Abstract

The invention provides a motor imagery electroencephalogram signal classification method, system and device based on multi-Riemannian kernel fusion features and a medium, and relates to the technical field of electroencephalogram signal classification, comprising: constructing and training a multi-Riemannian kernel fusion feature classification model for a subject; to-be-classified MI-EEG signals of a subject are collected and preprocessed, and optimized MI-EEG signals are obtained; performing multi-Riemannian kernel feature extraction on each section of optimized MI-EEG signal to obtain three kernel feature matrixes, performing centralization, dimension reduction and normalization processing in sequence, and performing feature fusion to obtain a fused feature matrix; and inputting the fusion kernel feature matrix into a multi-Riemann kernel fusion feature classification model to carry out category prediction so as to obtain a category label of the MI-EEG signal. According to the method, the kernel features of the MI-EEG signals on the multiple manifolds are fused, meanwhile, the influence of body difference is fully considered, richer and more comprehensive feature information is provided for classification, and the interpretability, accuracy and reliability of classification are improved.

Description

Background Art

[0002] Motor Imagery (MI) refers to the behavior of an individual imagining a specific movement (such as imagining the movement of the left or right hand) without actually performing the movement. This phenomenon has received extensive attention in fields such as neuroscience. Electroencephalogram (EEG) is a medical imaging technology that can read the scalp electrical signals generated by brain activities, and has advantages such as non-invasiveness, high temporal resolution, and low cost. The Motor Imagery Electroencephalogram (MI-EEG) signals collected using EEG, which are generated by motor imagery, are non-invasive and have high temporal resolution, and are the core research objects in the field of Brain-Computer Interface (BCI). The BCI technology based on MI-EEG signals can convert the user's movement intention into control instructions, thereby helping patients with limb dysfunction to control exoskeletons, wheelchairs, or virtual devices, significantly improving their quality of life. In addition, it also has important value in neurorehabilitation training, human-computer interaction optimization, and cognitive science research. However, due to the characteristics of MI-EEG signals such as non-stationarity, low signal-to-noise ratio, and large individual differences, how to efficiently extract its features and achieve high-precision classification has always been a technical difficulty in this field.

[0003] In the prior art, the processing methods of MI-EEG signals include analysis methods based on Euclidean space, deep learning analysis methods using Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN), and analysis methods based on Riemannian space. Among them, the feature extraction and classification methods based on Euclidean space are difficult to accurately depict the true features of brain activities and are difficult to effectively handle the complexity and non-Euclidean characteristics of signals; deep learning analysis methods often lack interpretability and have high requirements for the amount of data. Since the acquisition cost of MI-EEG signals is relatively high and the differences between different subjects are relatively large, it is difficult to obtain sufficient and representative data, and the application is greatly limited; the analysis method of MI-EEG signals based on Riemannian space takes the covariance matrix of MI-EEG signals as the conversion symbol from Euclidean space to Riemannian space, making it located on the Symmetric Positive Definite (SPD) manifold, and uses its inherent geometric characteristics to improve the performance of the classifier. This geometric method can not only effectively handle the non-Euclidean characteristics of MI-EEG signals, process the nonlinearity and non-stationarity of data, but also reduce the dependence on a large amount of data and improve the interpretability of the model, and to a certain extent overcomes the limitations of the methods based on Euclidean space and deep learning in processing MI-EEG signals.

[0004] However, the current application of Riemannian space in MI-EEG signal processing is usually limited to the SPD manifold, mainly focusing on the covariance matrix change and the calculation of geometric metrics. Although this method can provide a certain geometric structure to analyze MI-EEG signals and help reveal the correlations and patterns in the signals to a certain extent, the geometric structure it relies on is relatively single and it is difficult to comprehensively capture the complex geometric characteristics of MI-EEG signals. Summary of the Invention

[0005] Aiming at the technical problem in the prior art that the MI-EEG signal analysis method based on Riemannian space uses a single SPD manifold for analysis, with a relatively single geometric structure and difficulty in comprehensively capturing the complex geometric characteristics of MI-EEG signals, the invention provides a method, system, device and medium for classifying motor imagery electroencephalogram signals based on multi-Riemannian kernel fusion features, which fuse the kernel features of MI-EEG signals on multiple manifolds and fully consider the influence of individual differences at the same time, providing richer and more comprehensive feature information for classification and effectively improving the interpretability, accuracy and reliability of MI-EEG classification.

[0006] In the first aspect, the invention provides a method for classifying motor imagery electroencephalogram signals based on multi-Riemannian kernel fusion features, and the steps include: S1. Construct and train a multi-Riemannian kernel fusion feature classification model for the subject; S2. Signal preprocessing: Collect the MI-EEG signals to be classified of the subject and perform preprocessing to obtain the corresponding optimized MI-EEG signals. The preprocessing includes filtering and noise reduction processing; S3. Multi-Riemannian kernel feature extraction: Perform multi-Riemannian kernel feature extraction on each segment of the optimized MI-EEG signal. The multi-Riemannian kernel feature extraction includes extracting the kernel feature matrices of the optimized MI-EEG signal on the SPD manifold and the Gaussian SPD manifold, and extracting the kernel feature matrix of the optimized MI-EEG signal on the Grassmann manifold under the condition of the optimal linear subspace dimension; S4. Multi-Riemannian kernel feature fusion: Centering, dimension reduction and normalization processing are sequentially performed on the kernel feature matrices of each segment of the optimized MI-EEG signal on the SPD manifold, the Gaussian SPD manifold and the Grassmann manifold, and the three processed kernel feature matrices are subjected to feature fusion to obtain a fused kernel feature matrix; S5. Feature classification: The fused kernel feature matrix is sent to the multi-Riemannian kernel fusion feature classification model for class prediction to obtain the class label of the MI-EEG signal.

[0007] It should be further noted that in step S1, the classifier of the multi-Riemann kernel fusion feature classification model is selected from one of the support vector machine (SVM), logistic regression classifier, random forest classifier, naive bayes classifier, multi-layer perceptron (MLP), and convolutional neural network (CNN).

[0008] It should be further noted that in step S1, the steps for obtaining the multi-Riemann kernel fusion feature classification model include: S101. Preprocess several segments of known-labeled subjects' MI-EEG signals to obtain corresponding optimized training MI-EEG signals. The preprocessing includes filtering and noise reduction. S102. Extract multi-Riemann kernel features from each segment of the optimized training MI-EEG signal. The multi-Riemann kernel feature extraction includes extracting the kernel feature matrices of the optimized training MI-EEG signal on the SPD manifold and the Gaussian SPD manifold, and traversing all possible linear subspace dimension values. For each linear subspace dimension value, extract the kernel feature matrix of the optimized training MI-EEG signal on the Grassmann manifold. S103. Centering, dimensionality reduction, and normalization are performed on the kernel feature matrices of each segment of the optimized training MI-EEG signal on the SPD manifold, the Gaussian SPD manifold, and the Grassmann manifold under different linear subspace dimensions in sequence. Feature fusion is performed on the three processed kernel feature matrices to obtain the fused kernel feature matrix of the optimized training MI-EEG signal under different linear subspace dimensions. S104. The fused kernel feature matrices of each segment of the optimized training MI-EEG signal under different linear subspace dimensions and the corresponding known labels are jointly used for classifier training to obtain the best model of the classifier and the linear subspace dimension on the Grassmann manifold with the highest accuracy. 。

[0009] It should be further noted that the classifier in step S1 is a support vector machine. The specific operation of step S104 is as follows: S1041. Perform k-fold cross-validation on the fused kernel feature matrix data corresponding to each linear subspace dimension. Randomly and evenly divide the fused kernel feature matrix data corresponding to one linear subspace dimension into k non-overlapping subsets D1, D2, …, D k ; S1042. Perform k iterations. In the i-th iteration: Select D iUsing the remaining k - 1 subsets as the training set and the other subset as the test set, train the SVM model with the training set. During the training process, use grid search or random search to adjust the regularization parameter C of the SVM model and the γ parameter of the rbf kernel function to minimize the training error; Use the trained SVM model to predict the samples in the test set D i Compare the prediction results with the known labels corresponding to each fused kernel feature matrix in the test set D i and calculate the classification accuracy ACG of this iteration i . The calculation formula for the classification accuracy is:

[0010] where i = 1, 2,..., k; S1043. Aggregate the classification accuracies obtained from k iterations and calculate the average value to obtain the k - fold average classification accuracy of this linear subspace dimension; S1044. Aggregate and compare the k - fold average classification accuracies of different linear subspace dimensions, and select the linear subspace dimension with the highest k - fold average classification accuracy as the best linear subspace dimension , and the corresponding SVM model is the best multi - Riemannian kernel fusion feature classification model.

[0011] It should be further noted that k = 5 - 10.

[0012] It should be further noted that in step S2, Butterworth band - pass filtering is used for filtering and wavelet denoising is used for noise reduction.

[0013] It should be further noted that when using Butterworth band - pass filtering for filtering, the frequency components between 8 - 30 Hz are retained.

[0014] It should be further noted that the specific operations of step S4 include: S401. Centralize the kernel feature matrix using the centralization formula. The centralization formula is:

[0015] In the formula, Kx Take KS, KG, KGr(q) respectively; KS is the SPD manifold kernel feature matrix; KG is the Gaussian SPD manifold kernel feature matrix; KGr(q) is the Grassmann manifold kernel feature matrix when the value of the linear subspace dimension is q ; N is the number of segments of the MI-EEG signal; is a matrix with dimension N×N and all element values being 1 / N; After centering processing, the centered kernel feature matrix is obtained respectively ; S402. Perform dimensionality reduction processing on the centered kernel feature matrix, that is, perform eigenvalue decomposition, extract the eigenvectors corresponding to the first m eigenvalues, and retain the real part of the eigenvectors to obtain 3 dimensionality-reduced kernel feature matrices , the dimension of each dimensionality-reduced kernel feature matrix is N×j, each row of the dimensionality-reduced kernel feature matrix represents a segment of MI-EEG signal, and each column represents a feature; S403. Perform L2 norm normalization processing on the dimensionality-reduced kernel feature matrix according to the row eigenvectors. Let a certain row eigenvector be , and the corresponding L2 norm normalization formula is:

[0016] Perform norm normalization processing on each row eigenvector in the dimensionality-reduced kernel feature matrix according to the L2 norm normalization formula to obtain the normalized kernel feature matrix ; S404. Concatenate and fuse the normalized kernel feature matrix row by row to obtain a fused kernel feature matrix with dimension N×3k. The expression of the fused kernel feature matrix is: .

[0017] In the second aspect, the present invention provides a motor imagery electroencephalogram signal classification system based on multi-Riemann kernel fusion features for implementing the above-mentioned motor imagery electroencephalogram signal classification method based on multi-Riemann kernel fusion features, including: A model construction and training module for constructing and training a multi-Riemann kernel fusion feature classification model; A signal preprocessing module for preprocessing the MI-EEG signal to be classified to obtain the corresponding optimized MI-EEG signal; A multi-Riemann kernel feature extraction module for extracting multi-Riemann kernel features from each segment of the optimized MI-EEG signal to obtain the kernel feature matrices of the optimized MI-EEG signal on the SPD manifold, Gaussian SPD manifold, and Grassmann manifold; A multi-Riemann kernel feature fusion module for sequentially performing centering, dimensionality reduction, and normalization processing on the kernel feature matrices of each segment of the optimized MI-EEG signal on the SPD manifold, Gaussian SPD manifold, and Grassmann manifold, and performing feature fusion on the three processed kernel feature matrices to obtain a fused kernel feature matrix; A feature classification module, configured to send the fused kernel feature matrix to a multi-Riemannian kernel fusion feature classification model for class prediction, so as to obtain the class label of the MI-EEG signal.

[0018] In a third aspect, the present invention provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. The processor is configured to implement the steps of the above-mentioned method for classifying motor imagery electroencephalogram signals based on multi-Riemannian kernel fusion features when executing the computer program.

[0019] In a fourth aspect, the present invention provides a storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned method for classifying motor imagery electroencephalogram signals based on multi-Riemannian kernel fusion features are implemented.

[0020] The beneficial effects of the present invention are as follows: 1. For the method for classifying motor imagery electroencephalogram signals based on multi-Riemannian kernel fusion features provided by the present invention, after preprocessing the MI-EEG signal, the kernel feature matrices of the optimized MI-EEG signal on the SPD manifold and the Gaussian SPD manifold are respectively extracted, and the kernel feature matrix of the optimized MI-EEG signal on the Grassmann manifold is extracted under the condition of the optimal linear subspace dimension. The three kernel feature matrices are sequentially subjected to centering, dimensionality reduction, and normalization processing. After feature fusion of the processed three kernel feature matrices, they are sent to a multi-Riemannian kernel fusion feature classification model for class prediction. Compared with single-manifold or simple fusion methods, the present invention can comprehensively capture the complex geometric characteristics of MI-EEG signals from different geometric perspectives by fusing kernel features on multiple manifolds, effectively avoiding information redundancy and loss, supplementing the deficiencies of SPD manifold analysis, providing richer and more comprehensive feature information for classification, significantly improving the feature expression ability, and enhancing the accuracy and reliability of classification.

[0021] 2. The present invention uses the Riemannian kernel function to extract features from the MI-EEG signal, sequentially performs centering, dimensionality reduction, and normalization processing on the extracted kernel feature matrices, and performs feature fusion operations on the processed three kernel feature matrices. Among them, using the Riemannian kernel function to extract features deeply explores the relationship between MI-EEG signals, provides a deeper perspective for the analysis of MI-EEG signals, and further extracts more discriminative features; the subsequent processing process has clear mathematical principles and physical meanings, the operation steps are clear, and the logic is clear, making the entire classification process highly interpretable, and it is possible to clearly understand the impact of each step on the classification result, which is convenient for optimizing and improving the model.

[0022] 3. The present invention constructs and trains a multi-Riemannian kernel fusion feature classification model for a subject. During the process of training the multi-Riemannian kernel fusion feature classification model, by traversing all possible linear subspace dimensions, combining k-fold cross-validation with the adjustment of classifier parameters, the optimal Grassmann manifold linear subspace dimension and the optimal multi-Riemannian kernel fusion feature classification model are determined for each subject, fully considering the influence of individual differences on MI-EEG signals, better capturing the low-dimensional geometric characteristics of MI-EEG signals, overcoming the blindness of traditional empirical parameter tuning, ensuring that the model can maintain optimal performance on different subjects or datasets, and enabling the model to more accurately reflect the internal structure of the data.

[0023] 4. In the process of preprocessing MI-EEG signals, the present invention adopts a preprocessing scheme combining Butterworth band-pass filtering (8 - 30 Hz) and wavelet denoising, precisely retaining the μ rhythm (8 - 12 Hz) and β rhythm (18 - 26 Hz) signals related to motor imagery, while effectively suppressing high-frequency noise and baseline drift. The standardized preprocessing process significantly improves the signal quality, laying a reliable foundation for subsequent feature extraction and classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] To more clearly illustrate the technical solutions of the present invention, the drawings required for description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0025] Figure 1 is a flowchart of a method for classifying motor imagery electroencephalogram signals based on multi-Riemannian kernel fusion features in an embodiment of the present invention.

[0026] Figure 2 is a schematic block diagram of a system for classifying motor imagery electroencephalogram signals based on multi-Riemannian kernel fusion features in an embodiment of the present invention.

[0027] Figure 3 is a schematic diagram of the hardware structure of an electronic device in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0028] To make the objectives, features, and advantages of the present invention more obvious and understandable, the technical solutions in the present invention will be clearly and completely described below in conjunction with the drawings in the specific embodiments of the present invention. Obviously, the embodiments described below are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0029] The method for classifying motor imagery electroencephalogram signals based on multi-Riemannian kernel fusion features involved in this application mainly targets the technical field of electroencephalogram signal classification. The technical solution includes constructing and training a multi-Riemannian kernel fusion feature classification model for the subject; collecting the MI-EEG signals to be classified of the subject and performing preprocessing to obtain the corresponding optimized MI-EEG signals; extracting multi-Riemannian kernel features for each segment of the optimized MI-EEG signals, and successively performing centering, dimensionality reduction, and normalization processing. Feature fusion is performed on the three processed kernel feature matrices to obtain a fused feature matrix; the fused kernel feature matrix is sent to the multi-Riemannian kernel fusion feature classification model for class prediction to obtain the class label of the MI-EEG signal. Compared with the prior art, after preprocessing the MI-EEG signals, this invention respectively extracts the kernel feature matrices of the optimized MI-EEG signals on the SPD manifold and the Gaussian SPD manifold, as well as the kernel feature matrix of the optimized MI-EEG signals on the Grassmann manifold under the condition of the optimal linear subspace dimension. The three kernel feature matrices are successively subjected to centering, dimensionality reduction, and normalization processing. After feature fusion of the three processed kernel feature matrices, they are sent to the multi-Riemannian kernel fusion feature classification model for class prediction.Compared with single manifold or simple fusion methods, the present invention can comprehensively capture the complex geometric characteristics of MI-EEG signals from different geometric perspectives by fusing kernel features on multiple manifolds, effectively avoiding information redundancy and loss, complementing the deficiencies of SPD manifold analysis, providing richer and more comprehensive feature information for classification, significantly enhancing the feature expression ability, and improving the accuracy and reliability of classification; after feature extraction of MI-EEG signals, centering, dimensionality reduction, and normalization processing are performed in sequence, and feature fusion operations are performed on the three processed kernel feature matrices. The Riemann kernel function is used for feature extraction to deeply explore the relationships between MI-EEG signals, providing a deeper perspective for the analysis of MI-EEG signals and further extracting more discriminative features; the subsequent processing processes have clear mathematical principles and physical meanings, with clear operation steps and clear logic, making the entire classification process highly interpretable, enabling a clear understanding of the impact of each step on the classification result and facilitating the optimization and improvement of the model; a multi-Riemann kernel fusion feature classification model is constructed and trained for the subjects. During the process of training the multi-Riemann kernel fusion feature classification model, by traversing all possible linear subspace dimensions, combined with k-fold cross-validation and adjustment of classifier parameters, the optimal Grassmann manifold linear subspace dimension and the optimal multi-Riemann kernel fusion feature classification model are determined for each subject, fully considering the influence of individual differences on MI-EEG signals, better capturing the low-dimensional geometric characteristics of MI-EEG signals, overcoming the blindness of traditional empirical parameter tuning, ensuring that the model can maintain optimal performance on different subjects or datasets, and enabling the model to more accurately reflect the internal structure of the data; during the preprocessing of MI-EEG signals, a preprocessing scheme combining Butterworth band-pass filtering (8-30 Hz) and wavelet denoising is adopted to accurately retain the mu rhythm (8-12 Hz) and beta rhythm (18-26 Hz) signals related to motor imagery, while effectively suppressing high-frequency noise and baseline drift. The standardized preprocessing process significantly improves the signal quality, laying a reliable foundation for subsequent feature extraction and classification.

[0030] The method for classifying motor imagery electroencephalogram signals based on multi-Riemann kernel fusion features involved in this application mainly addresses the technical problem that the existing Riemann space-based analysis methods for MI-EEG signals use a single SPD manifold for analysis, with a relatively simple geometric structure and difficulty in comprehensively capturing the complex geometric characteristics of MI-EEG signals.

[0031] The method for classifying motor imagery electroencephalogram signals based on multi-Riemann kernel fusion features involved in this application will be described in detail below. For the purpose of illustration rather than limitation, specific details such as specific system structures and technologies are proposed to thoroughly understand the embodiments of this application. However, those skilled in the art should clearly understand that this application can also be implemented in other embodiments without these specific details.

[0032] In the method for classifying motor imagery electroencephalogram signals based on multi-Riemann kernel fusion features involved in this application, the term "including" indicates the presence of the described features, wholes, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components, and / or their sets. The terms "including", "comprising", "having", and their variants all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0033] For the convenience of clearly describing the technical solutions of this application, words such as "first" and "second" are used to distinguish the same items or similar items with basically the same functions and roles. Those skilled in the art can understand that words such as "first" and "second" do not limit the quantity and execution order, and words such as "first" and "second" do not necessarily limit being different.

[0034] The statements such as "an embodiment" or "some embodiments" described in this application mean that the specific features, structures, or characteristics described in the embodiment are included in one or more embodiments of this application. Thus, statements such as "in an embodiment", "in some embodiments", "in other some embodiments", "in still other embodiments" that appear in different places in this application do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways.

[0035] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

[0036] The method for classifying motor imagery electroencephalogram signals based on multi-Riemann kernel fusion features provided by the embodiments of the present invention is executed by a computer device. Correspondingly, the system for classifying motor imagery electroencephalogram signals based on multi-Riemann kernel fusion features runs in the computer device.

[0037] Figure 1 is a flowchart of the method for classifying motor imagery electroencephalogram signals based on multi-Riemann kernel fusion features in an embodiment of the present invention. Among them, Figure 1 The execution subject can be a system for classifying motor imagery electroencephalogram signals based on multi-Riemann kernel fusion features. According to different requirements, the order of the steps in this flowchart can be changed, and some can be omitted.

[0038] Such as Figure 1As shown, the method for classifying motor imagery EEG signals based on multi-Riemannian kernel fusion features includes: Step S1: Construct and train a multi-Riemannian kernel fusion feature classification model for the subject.

[0039] The construction and training of a customized multi-Riemannian kernel fusion feature classification model can fully consider the individual differences of the subject, enabling the model to better adapt to the EEG signal characteristics of different subjects, and improving the classification accuracy and generalization ability.

[0040] In some specific embodiments, the classifier of the multi-Riemannian kernel fusion feature classification model is selected from one of support vector machine (SVM), logistic regression classifier, random forest classifier, naive Bayes classifier, multi-layer perceptron (MLP), and convolutional neural network (CNN).

[0041] Different classifiers have their unique advantages and applicable scenarios, and can be flexibly selected according to the specific characteristics of EEG signal data, computing resources, and actual application requirements to optimize the classification effect and improve the adaptability and practicality of the model.

[0042] In some specific embodiments, the steps for obtaining the multi-Riemannian kernel fusion feature classification model include: S101. Preprocess several segments of the subject's MI-EEG signals with known labels to obtain corresponding optimized training MI-EEG signals. The preprocessing includes filtering and noise reduction. S102. Extract multi-Riemannian kernel features from each segment of the optimized training MI-EEG signal. The multi-Riemannian kernel feature extraction includes extracting the kernel feature matrices of the optimized training MI-EEG signal on the SPD manifold and the Gaussian SPD manifold, and traversing all possible linear subspace dimension values. For each linear subspace dimension value, extract the kernel feature matrix of the optimized training MI-EEG signal on the Grassmann manifold. S103. Centering, dimensionality reduction, and normalization are sequentially performed on the kernel feature matrices of each segment of the optimized training MI-EEG signal on the SPD manifold, the Gaussian SPD manifold, and the Grassmann manifold under different linear subspace dimensions. The three processed kernel feature matrices are feature fused to obtain the fused kernel feature matrix of the optimized training MI-EEG signal under different linear subspace dimensions. S104. The fused kernel feature matrices of each segment of the optimized training MI-EEG signal under different linear subspace dimensions and the corresponding known labels are jointly used for classifier training to obtain the best model of the classifier and the linear subspace dimension on the Grassmann manifold with the highest accuracy.

[0043] Preprocessing the MI-EEG signals with known labels and then performing multi-Riemann kernel feature extraction can enrich the feature dimensions while ensuring the quality of training data; in the process of multi-Riemann kernel feature extraction, all possible linear subspace dimension values ​​are traversed, and the kernel feature matrix of the optimized training MI-EEG signal on the Grassmann manifold is extracted for each linear subspace dimension value to fully mine the data features; after feature fusion, the optimal model and linear subspace dimension are determined through classifier training, which can improve the generalization ability of the model, avoid overfitting, and enable the model to better adapt to the EEG signal classification tasks under different subjects and experimental conditions.

[0044] In some specific embodiments, the classifier in step S1 is a support vector machine, and the specific operation of step S104 is: S1041. Perform k-fold cross validation on the fusion kernel feature matrix data corresponding to each linear subspace dimension, and randomly and evenly divide the fusion kernel feature matrix data corresponding to a linear subspace dimension into k mutually non-overlapping subsets D1, D2, ..., D k ; S1042. Perform k iterations. In the i-th iteration: Select D i As the test set, the remaining k-1 subsets are combined as the training set, and the SVM model is trained using the training set. During the training process, grid search or random search is used to adjust the regularization parameter C of the SVM model and the γ parameter of the rbf kernel function to minimize the training error; Use the trained SVM model to test the set D i The samples in the prediction are predicted and the prediction results are compared with the test set D i Compare the known labels corresponding to each fusion kernel feature matrix in the fusion kernel feature matrix to calculate the classification accuracy of this iteration ACG i , the calculation formula of classification accuracy is:

[0045] Where i=1,2,…,k; S1043. Summarize the classification accuracy obtained from k iterations and calculate the average value to obtain the k-fold average classification accuracy of the linear subspace dimension; S1044. Summarize and compare the k-fold average classification accuracy of different linear subspace dimensions, and select the linear subspace dimension with the highest k-fold average classification accuracy as the optimal linear subspace dimension , the corresponding SVM model is the best multi-Riemann kernel fusion feature classification model.

[0046] For the support vector machine classifier, k-fold cross-validation combined with grid search or random search is used to adjust the parameters. K-fold cross-validation can make full use of limited data, evaluate the performance of the model on different data subsets, and reduce the bias and variance of model evaluation. During the training process, the regularization parameter C and the γ parameter of the rbf kernel function are adjusted, which can optimize the model complexity and fitting ability, find the best parameter combination, and thus significantly improve the classification accuracy and stability of the support vector machine classifier.

[0047] In some specific embodiments, k = 5 - 10.

[0048] Step S2: Collect the MI-EEG signals to be classified of the subjects and perform preprocessing to obtain the corresponding optimized MI-EEG signals. The preprocessing includes filtering and noise reduction.

[0049] The preprocessing can remove the noise and interference in the MI-EEG signals to be classified, improve the signal quality, lay a foundation for accurately extracting features and classification subsequently, and reduce the possibility of misclassification.

[0050] In some specific embodiments, Butterworth band-pass filtering is used for filtering, and wavelet denoising is used for noise reduction.

[0051] The Butterworth filter is a filtering technology widely used in the preprocessing of MI-EEG signals. Its main purpose is to remove high-frequency noise and alternating current interference in the signals. A significant feature of this filter is that within the passband, its frequency response curve is relatively flat without obvious fluctuations, while in the stopband, the response gradually drops to zero. This characteristic enables the Butterworth filter to effectively retain the signals within the passband while suppressing the interference or noise within the stopband, and thus can accurately retain the frequency components closely related to motor imagery and enhance the effective signals. The principle of wavelet denoising is that after the signal is decomposed by wavelets, the wavelet coefficients of the useful signals are larger, and the wavelet coefficients of the noise are smaller, that is, the wavelet coefficients of the noise are smaller than those of the signals. By selecting a suitable threshold, the wavelet coefficients greater than the threshold are considered to be generated by signals and should be retained, while those less than the threshold are considered to be generated by noise and are set to zero to achieve the purpose of denoising. Therefore, wavelet denoising can effectively remove high-frequency noise and baseline drift and reduce noise interference. The combination of the two significantly improves the quality of the MI-EEG signals and provides a clean and stable signal basis for accurately extracting features and reliable classification subsequently.

[0052] In some specific embodiments, when using Butterworth band-pass filtering for filtering, the frequency components between 8 - 30 Hz are retained.

[0053] The 8 - 30Hz frequency component range is the key frequency band of motor imagery EEG signals. Among them, the μ rhythm (8 - 12Hz) and β rhythm (18 - 26Hz) play important indicative roles in motor imagery tasks. Precisely retaining these frequency bands can maximize the retention of information related to motor imagery, reduce the interference of other irrelevant frequency components, improve the accuracy of subsequent feature extraction and classification, and make the classification results better reflect the true motor imagery intention.

[0054] Step S3: Perform multi - Riemannian kernel feature extraction on each segment of the optimized MI - EEG signal. The multi - Riemannian kernel feature extraction includes extracting the kernel feature matrices of the optimized MI - EEG signal on the SPD manifold and Gaussian SPD manifold, and extracting the kernel feature matrix of the optimized MI - EEG signal on the Grassmann manifold under the condition of the optimal linear subspace dimension.

[0055] Obtaining the kernel feature matrices of the optimized MI - EEG signal from multiple different manifold perspectives comprehensively enriches the feature information, avoids information loss caused by single - feature extraction, and enhances the representativeness and discriminability of the features.

[0056] Step S4: Centering, dimensionality reduction, and normalization processing are successively performed on the kernel feature matrices of each segment of the optimized MI - EEG signal on the SPD manifold, Gaussian SPD manifold, and Grassmann manifold. Feature fusion is performed on the three processed kernel feature matrices to obtain a fused kernel feature matrix.

[0057] Successively performing centering, dimensionality reduction, and normalization processing on the three types of kernel feature matrices and performing feature fusion operations on the three processed kernel feature matrices can eliminate the bias in the feature matrices, reduce the data dimension and calculation amount, unify the data scale, and fuse multi - manifold features, making the final feature matrix more valuable for classification and improving the classification performance.

[0058] In some specific embodiments, the specific operations of step S4 include: S401. Perform centering processing on the kernel feature matrix using the centering formula. The centering formula is:

[0059] where Kx respectively take KS, KG, KGr(q) ; KS is the SPD manifold kernel feature matrix; KG is the Gaussian SPD manifold kernel feature matrix; KGr(q) is the Grassmann manifold kernel feature matrix when the value of the linear subspace dimension is q ; N is the number of segments of the MI-EEG signal; is a matrix with dimension N×N and all element values ​​are 1 / N; After centralization, the kernel feature matrices after centralization are obtained: ; S402. Perform dimensionality reduction processing on the core feature matrix after centralization, that is, perform eigenvalue decomposition, extract the eigenvectors corresponding to the first m eigenvalues, and retain the real part of the eigenvector to obtain three dimensionality reduction core feature matrices , the dimension of each dimensionality reduction kernel feature matrix is ​​N×j, each row of the dimensionality reduction kernel feature matrix represents a segment of MI-EEG signal, and each column represents a feature; S403. Perform L2 norm normalization on the reduced-dimensional kernel feature matrix according to the row feature vector. Suppose a row feature vector is , the corresponding L2 norm normalization formula is:

[0060] Each row of feature vectors in the dimension reduction kernel feature matrix is ​​normalized according to the L2 norm normalization formula to obtain the normalized kernel feature matrix ; S404. Normalize the kernel feature matrix The concatenation and fusion are performed row by row to obtain a fusion kernel feature matrix with a dimension of N×3k. The expression of the fusion kernel feature matrix is: .

[0061] The centering operation can remove the deviation of the feature matrix and make the data distribution more concentrated; the dimensionality reduction operation reduces the data dimension through eigenvalue decomposition, reduces the computational complexity and removes redundant information; the normalization operation can unify the data scale and improve the efficiency and stability of model training; the three sets of manifold kernel feature matrices are spliced ​​and fused by row to form a fused kernel feature matrix, which can comprehensively reflect the characteristics of EEG signals, enhance the classification model's ability to distinguish different motor imagery EEG signals, and improve classification accuracy.

[0062] Step S5, transmitting the fusion kernel feature matrix to the multi-Riemann kernel fusion feature classification model for category prediction to obtain the category label of the MI-EEG signal.

[0063] By using the trained model and optimized feature matrix, accurate classification of EEG signals can be achieved and reliable category labels can be output, which can provide effective support for practical applications such as brain-computer interface control.

[0064] In a specific embodiment, a method for classifying motor imagery EEG signals based on multi-Riemann kernel fusion features includes: Step S1. Construct and train a multi-Riemannian kernel fusion feature classification model for the subject; Among them, the classifier of the multi-Riemannian kernel fusion feature classification model is a support vector machine (SVM). The steps to obtain the multi-Riemannian kernel fusion feature classification model include: S101. Preprocess several segments of the subject's MI-EEG signals with known labels to obtain the corresponding optimized training MI-EEG signals. The preprocessing includes filtering using a Butterworth band-pass filter and denoising using the wavelet denoising method. When using the Butterworth band-pass filter for filtering, a 5th-order Butterworth filter is used to filter the MI-EEG signals, and the frequency components between 8 - 30 Hz are retained; Among them, the magnitude-squared function of the Butterworth filter is:

[0065] In the formula, H represents the transfer function of the Butterworth filter; j is the imaginary unit, satisfying j 2 = -1; is the angular frequency variable, representing the frequency of the input signal; are the band-pass cut-off frequencies; is the order of the Butterworth filter; The wavelet function used for wavelet transform is set to 'db4', that is, the vanishing moment is 4; the threshold size is 0.2; S102. Extract multi-Riemannian kernel features from each segment of the optimized training MI-EEG signals. The multi-Riemannian kernel feature extraction includes extracting the kernel feature matrices of the optimized training MI-EEG signals on the SPD manifold and the Gaussian SPD manifold, and traversing all possible linear subspace dimension values, and respectively extracting the kernel feature matrices of the optimized training MI-EEG signals on the Grassmann manifold for each linear subspace dimension value; Among them, the steps to extract the kernel feature matrix of the optimized training MI-EEG signal on the SPD manifold include: S1021.1. Using the sample covariance matrix, convert the original MI-EEG signal into a d-dimensional SPD matrix. For the MI-EEG signal , its sample covariance matrix is expressed as:

[0066] In the formula, represents the -th segment of the subject's MI-EEG signal, ; Denotes the number of MI-EEG channels; Denotes the number of sampling points in each segment of MI-EEG signal; Denotes at moment, the MI-EEG signals from all channels, ; is 's mean vector, ; is a symmetric positive definite matrix, spanning the SPD manifold ; S1021.2. Suppose a subject has a total of N segments of MI-EEG signals, that is, N sample covariance matrices are obtained. For any two elements on the SPD manifold, its SPD manifold kernel function is defined as:

[0067] Calculate the kernel function values between pairwise covariance matrices according to the SPD manifold kernel function formula, and then obtain the SPD manifold kernel feature matrix , Denoted as: ; The steps to extract the kernel feature matrix of the optimized training MI-EEG signal in the Gaussian SPD manifold include: S1022.1. Model each segment of MI-EEG signal as a single Gaussian distribution. Given the MI-EEG signal , define its Gaussian distribution as , where and respectively represent 's mean and covariance.

[0068] Embed the Gaussian distribution into -dimensional SPD space , -dimensional SPD matrix can be uniquely represented as the Gaussian model , and the Gaussian model modeling formula is:

[0069] where, is decomposed by , ; For each segment of MI-EEG signal , calculate its mean vector and covariance matrix , modeled using the Gaussian model formula to obtain -dimensional SPD matrix , which resides in another SPD manifold . The MI-EEG signal is transformed from the spatio-temporal domain description in the traditional Euclidean space to the representation on the Gaussian SPD manifold; S1022.2. For any two elements on the Gaussian SPD manifold , its Gaussian SPD manifold kernel function is expressed as:

[0070] For each pair of matrices and , calculate the kernel value between them using the Gaussian SPD manifold kernel function formula , and then construct -dimensional kernel feature matrix , which contains the similarities between all sample pairs, expressed as: ; The steps to extract the kernel feature matrix of the optimized training MI-EEG signal on the Grassmann manifold include: S1023.1. For each segment of MI-EEG signal , by performing eigen-decomposition on to obtain its -dimensional linear subspace , the formula is as follows:

[0071] In the formula, and respectively represent the first largest eigenvalues and their corresponding eigenvectors; After obtaining , further obtain the element representation in the Grassmann manifold, that is, the projection matrix . The required -dimensional linear subspace resides in the Grassmann manifold . Therefore, the MI-EEG signal can be transformed from the spatio-temporal domain description in the traditional Euclidean space to the representation on the Grassmann manifold through the projection matrix ; S1023.2. For any two elements on , , and its Grassmann manifold kernel function is expressed as:

[0072] Using the Grassmann manifold kernel function formula, calculate the kernel function values between different projection matrices , and then obtain -dimensional Grassmann kernel feature matrix , which is expressed as: ; S103. For each segment of the optimized training MI-EEG signal, perform centering, dimensionality reduction, and normalization on the kernel feature matrices on the SPD manifold, Gaussian SPD manifold, and Grassmann manifold with different linear subspace dimensions in sequence, and perform feature fusion on the three processed kernel feature matrices to obtain the fused kernel feature matrix of the optimized training MI-EEG signal under different linear subspace dimensions; The specific operations of step S103 include: S1031. Use the centering formula to perform centering processing on the kernel feature matrix. The centering formula is:

[0073] In the formula, Kx respectively take KS, KG, KGr(q) ; KS is the SPD manifold kernel feature matrix; KG is the Gaussian SPD manifold kernel feature matrix; KGr(q) is the Grassmann manifold kernel feature matrix when the value of the linear subspace dimension is q ; N is the number of segments of the MI-EEG signal; is a matrix with dimension N×N and all element values equal to 1 / N; After centering processing, the centered kernel feature matrices are obtained respectively; S1032. Perform dimensionality reduction processing on the centered kernel feature matrix, that is, perform eigenvalue decomposition, extract the eigenvectors corresponding to the first m eigenvalues, and retain the real part of the eigenvectors to obtain 3 dimensionality-reduced kernel feature matrices , and the dimension of each dimensionality-reduced kernel feature matrix is N×j. Each row of the dimensionality-reduced kernel feature matrix represents a segment of the MI-EEG signal, and each column represents a feature; S1033. Perform L2 norm normalization on the reduced-dimensional kernel feature matrix according to the row feature vector. Suppose a row feature vector is , the corresponding L2 norm normalization formula is:

[0074] Each row of feature vectors in the dimension reduction kernel feature matrix is ​​normalized according to the L2 norm normalization formula to obtain the normalized kernel feature matrix ; S1034. Normalize the kernel feature matrix The concatenation and fusion are performed row by row to obtain a fusion kernel feature matrix with a dimension of N×3k. The expression of the fusion kernel feature matrix is: ; S104. The fusion kernel feature matrix of each optimized training MI-EEG signal in different linear subspace dimensions is used together with the corresponding known labels to train the classifier, and the best classifier model and the linear subspace dimension on the Grassmann manifold with the highest accuracy are obtained. ; The specific operations of step S104 are: S1041. Perform k-fold cross validation on the fusion kernel feature matrix data corresponding to each linear subspace dimension, and randomly and evenly divide the fusion kernel feature matrix data corresponding to a linear subspace dimension into k mutually non-overlapping subsets D1, D2, ..., D k ; S1042. Perform 5 iterations. In the i-th iteration: Select D i As the test set, the remaining 4 subsets are combined as the training set, and the SVM model is trained using the training set. During the training process, grid search or random search is used to adjust the regularization parameter C of the SVM model and the γ parameter of the rbf kernel function to minimize the training error; Use the trained SVM model to test the set D i The samples in the prediction are predicted and the prediction results are compared with the test set D i Compare the known labels corresponding to each fusion kernel feature matrix in the fusion kernel feature matrix to calculate the classification accuracy of this iteration ACG i , the calculation formula of classification accuracy is:

[0075] Where i=1,2,…,5; S1043. Summarize the classification accuracy rates obtained from the five iterations and calculate the average value to obtain the 5-fold average classification accuracy rate of the linear subspace dimension; S1044. Aggregate and compare the k-fold average classification accuracies of different linear subspace dimensions, and select the linear subspace dimension with the highest k-fold average classification accuracy as the optimal linear subspace dimension. The corresponding SVM model is the optimal multi-Riemannian kernel fusion feature classification model. Step S2: Collect the MI-EEG signals to be classified of the subjects and perform preprocessing to obtain the corresponding optimized MI-EEG signals. The preprocessing includes filtering using Butterworth band-pass filtering and noise reduction using wavelet denoising. When using Butterworth band-pass filtering for filtering, the frequency components between 8 - 30 Hz are retained. The preprocessing steps are the same as those in step S101. Step S3: Extract multi-Riemannian kernel features from each segment of the optimized MI-EEG signals. The multi-Riemannian kernel feature extraction includes extracting the kernel feature matrices of the optimized MI-EEG signals on the SPD manifold, Gaussian SPD manifold, and extracting the kernel feature matrix of the optimized MI-EEG signals on the Grassmann manifold under the condition of the optimal linear subspace dimension. Among them, the steps of extracting the kernel feature matrix of the optimized training MI-EEG signals on the SPD manifold are the same as those in steps S1021.1 - S1021.2, the steps of extracting the kernel feature matrix of the optimized training MI-EEG signals on the Gaussian SPD manifold are the same as those in steps S1022.1 - S1022.2, and the steps of extracting the kernel feature matrix of the optimized training MI-EEG signals on the Grassmann manifold are the same as those in steps S1023.1 - S1023.2. Step S4: Centering, dimensionality reduction, and normalization are sequentially performed on the kernel feature matrices of each segment of the optimized MI-EEG signals on the SPD manifold, Gaussian SPD manifold, and Grassmann manifold, and the three processed kernel feature matrices are subjected to feature fusion to obtain a fused kernel feature matrix. The specific operations of step S4 are the same as those in steps S1031 - S1034. Step S5: The fused kernel feature matrix is sent to the multi-Riemannian kernel fusion feature classification model for class prediction to obtain the class labels of the MI-EEG signals.

[0076] The following is an embodiment of the motor imagery electroencephalogram signal classification system based on multi-Riemannian kernel fusion features provided by the present disclosure. This active load shedding optimization system and the motor imagery electroencephalogram signal classification method based on multi-Riemannian kernel fusion features in the above embodiments belong to the same inventive concept. For the details not described in detail in the embodiment of the motor imagery electroencephalogram signal classification system based on multi-Riemannian kernel fusion features, reference can be made to the embodiments of the motor imagery electroencephalogram signal classification method based on multi-Riemannian kernel fusion features.

[0077] The mobile terminal implementing various embodiments of the present invention will now be described with reference to the accompanying drawings. In the following description, suffixes such as "module", "component", or "unit" used to denote elements are only for facilitating the description of the embodiments of the present invention, and have no specific meaning in themselves. Therefore, "module" and "component" can be used interchangeably.

[0078] As Figure 2 shown, the motor imagery electroencephalogram (MI-EEG) signal classification system based on multi-Riemannian kernel fusion features includes: A model construction and training module for constructing and training a multi-Riemannian kernel fusion feature classification model; A signal preprocessing module for preprocessing the MI-EEG signal to be classified to obtain the corresponding optimized MI-EEG signal; A multi-Riemannian kernel feature extraction module for extracting multi-Riemannian kernel features from each segment of the optimized MI-EEG signal to obtain the kernel feature matrices of the optimized MI-EEG signal on the SPD manifold, Gaussian SPD manifold, and Grassmann manifold; A multi-Riemannian kernel feature fusion module for successively performing centering, dimensionality reduction, and normalization on the kernel feature matrices of each segment of the optimized MI-EEG signal on the SPD manifold, Gaussian SPD manifold, and Grassmann manifold, and performing feature fusion on the three processed kernel feature matrices to obtain a fused kernel feature matrix; A feature classification module for sending the fused kernel feature matrix to the multi-Riemannian kernel fusion feature classification model for class prediction to obtain the class label of the MI-EEG signal.

[0079] The motor imagery electroencephalogram signal classification system of this embodiment is used to implement the motor imagery electroencephalogram signal classification method based on multi-Riemannian kernel fusion features, and the steps include: S1. Construct and train a multi-Riemannian kernel fusion feature classification model for the subject; S2. Signal preprocessing: Collect the MI-EEG signal to be classified of the subject and perform preprocessing to obtain the corresponding optimized MI-EEG signal. The preprocessing includes filtering and noise reduction; S3. Multi-Riemannian kernel feature extraction: Extract multi-Riemannian kernel features from each segment of the optimized MI-EEG signal. The multi-Riemannian kernel feature extraction includes extracting the kernel feature matrices of the optimized MI-EEG signal on the SPD manifold and Gaussian SPD manifold, and extracting the kernel feature matrix of the optimized MI-EEG signal on the Grassmann manifold under the condition of the optimal linear subspace dimension; S4. Multi-Riemannian kernel feature fusion: Centralize, reduce the dimension, and normalize the kernel feature matrices of each optimized MI-EEG signal on the SPD manifold, Gaussian SPD manifold, and Grassmann manifold in sequence, and perform feature fusion on the three processed kernel feature matrices to obtain a fused kernel feature matrix; S5. Feature classification: Send the fused kernel feature matrix to the multi-Riemannian kernel fusion feature classification model for class prediction to obtain the class label of the MI-EEG signal.

[0080] This application also provides an electronic device for implementing each embodiment of the present invention. The electronic device includes a memory, a processor, and a computer program stored on the memory and executable on the processor.

[0081] Those skilled in the art can understand that the structure of the electronic device involved in the embodiments of the present invention does not constitute a limitation on the electronic device. The electronic device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0082] Figure 3 Schematic diagram of the hardware structure of an electronic device for implementing each embodiment of the present invention.

[0083] The electronic device includes, but is not limited to, components such as a processor and a memory. Those skilled in the art can understand that the structure of the electronic device involved in the embodiments of the present invention does not constitute a limitation on the electronic device. The electronic device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0084] In the embodiments of the present invention, the electronic device includes, but is not limited to, a laptop computer, a desktop computer, a workbench, a personal digital assistant, a server, a blade server, a mainframe computer, and other suitable computers. The electronic device may also represent various forms of mobile devices and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the embodiments of the present application described herein and / or claimed.

[0085] In the embodiments of the present application, the processor may be implemented by using at least one of an application specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a processor, a controller, a microcontroller, a microprocessor, and an electronic unit designed to execute the functions described herein. In some cases, such an implementation may be implemented in the controller. For a software implementation, an implementation of a process or function may be implemented with a separate software module that allows execution of at least one function or operation. The software code may be implemented by a software application (or program) written in any appropriate programming language. The software code may be stored in a memory and executed by the controller.

[0086] In addition, the electronic device includes some functional modules not shown, which will not be elaborated here.

[0087] Those skilled in the art can understand that various aspects of the electronic device provided in the present application can be implemented as a system, a method, or a program product. Therefore, various aspects of the present disclosure can be specifically implemented in the following forms, namely: a complete hardware implementation, a complete software implementation (including firmware, microcode, etc.), or an implementation combining hardware and software aspects, which can be collectively referred to as "circuit", "module", or "system" here.

[0088] The present application also provides a storage medium in which a program product capable of implementing a method for classifying motor imagery electroencephalogram signals based on multi-Riemann kernel fusion features is stored. In some possible implementation manners, various aspects of the present disclosure can also be implemented in the form of a program product, which includes program code. When the program product runs on a terminal device, the program code is used to cause the terminal device to execute the steps according to various exemplary embodiments of the present disclosure described in the above "Exemplary Method" section of this specification.

[0089] The storage medium may employ any combination of one or more readable media. The readable media may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the foregoing. More specific examples (a non-exhaustive list) of the readable storage medium include: an electrical connection having one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0090] The foregoing description of the disclosed embodiments enables those skilled in the art to practice or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for classifying motor imagery EEG signals based on multi-Riemann kernel fusion features, characterized in that the steps include: S1. Construct and train a multi-Riemann kernel fusion feature classification model for the subjects; S2. collecting the subject's MI-EEG signal to be classified and preprocessing it to obtain the corresponding optimized MI-EEG signal, the preprocessing including filtering and noise reduction; S3. Perform multi-Riemann kernel feature extraction on each segment of optimized MI-EEG signal. Multi-Riemann kernel feature extraction includes extracting the kernel feature matrix of the optimized MI-EEG signal on the SPD manifold, Gaussian SPD manifold, and the optimal linear subspace dimension. Extract the kernel feature matrix of the optimized MI-EEG signal on the Grassmann manifold under certain conditions; S4. The kernel feature matrices of each optimized MI-EEG signal on the SPD manifold, Gaussian SPD manifold and Grassmann manifold are sequentially centered, dimensionally reduced and normalized, and the three processed kernel feature matrices are subjected to feature fusion to obtain a fused kernel feature matrix; S5. The fusion kernel feature matrix is ​​transmitted to the multi-Riemann kernel fusion feature classification model for category prediction to obtain the category label of the MI-EEG signal.

2. The motor imagery EEG signal classification method according to claim 1, characterized in that: In step S1, the classifier of the multi-Riemann kernel fusion feature classification model is selected from one of a support vector machine, a logistic regression classifier, a random forest classifier, a naive Bayes classifier, a multilayer perceptron, and a convolutional neural network.

3. The motor imagery EEG signal classification method according to claim 1, characterized in that: In step S1, the step of obtaining the multi-Riemann kernel fusion feature classification model includes: S101. Preprocessing several segments of MI-EEG signals of subjects with known labels to obtain corresponding optimized training MI-EEG signals, wherein the preprocessing includes filtering and noise reduction; S102. Perform multi-Riemann kernel feature extraction on each segment of optimized training MI-EEG signal. The multi-Riemann kernel feature extraction includes extracting the kernel feature matrix of the optimized training MI-EEG signal on the SPD manifold and the Gaussian SPD manifold, and traversing all possible linear subspace dimension values, and extracting the kernel feature matrix of the optimized training MI-EEG signal on the Grassmann manifold for each linear subspace dimension value; S103. The kernel feature matrices of each optimized training MI-EEG signal on the SPD manifold, Gaussian SPD manifold and Grassmann manifold under different linear subspace dimensions are sequentially centralized, dimensionally reduced and normalized, and the three processed kernel feature matrices are subjected to feature fusion to obtain the fused kernel feature matrices of the optimized training MI-EEG signal under different linear subspace dimensions; S104. The fusion kernel feature matrix of each optimized training MI-EEG signal in different linear subspace dimensions is used together with the corresponding known labels to train the classifier, and the best classifier model and the linear subspace dimension on the Grassmann manifold with the highest accuracy are obtained. .

4. The motor imagery EEG signal classification method according to claim 3, characterized in that: The classifier of step S1 is a support vector machine, and the specific operation of step S104 is: S1041. Perform k-fold cross validation on the fusion kernel feature matrix data corresponding to each linear subspace dimension, and randomly and evenly divide the fusion kernel feature matrix data corresponding to a linear subspace dimension into k mutually non-overlapping subsets D1, D2, ..., D k ; S1042. Perform k iterations. In the i-th iteration: Select D i As the test set, the remaining k-1 subsets are combined as the training set, and the SVM model is trained using the training set. During the training process, grid search or random search is used to adjust the regularization parameter C of the SVM model and the γ parameter of the rbf kernel function to minimize the training error; Use the trained SVM model to test the set D i The samples in the prediction are predicted and the prediction results are compared with the test set D i Compare the known labels corresponding to each fusion kernel feature matrix in the fusion kernel feature matrix to calculate the classification accuracy of this iteration ACG i , the calculation formula of classification accuracy is: Where i=1,2,…,k; S1043. Summarize the classification accuracy obtained from k iterations and calculate the average value to obtain the k-fold average classification accuracy of the linear subspace dimension; S1044. Summarize and compare the k-fold average classification accuracy of different linear subspace dimensions, and select the linear subspace dimension with the highest k-fold average classification accuracy as the optimal linear subspace dimension , the corresponding SVM model is the best multi-Riemann kernel fusion feature classification model.

5. The motor imagery EEG signal classification method according to claim 1, characterized in that: In step S2, Butterworth bandpass filtering is used for filtering, and wavelet denoising is used for noise reduction.

6. The motor imagery EEG signal classification method according to claim 5, characterized in that: When using Butterworth bandpass filtering for filtering, the frequency components between 8-30 Hz are retained.

7. The motor imagery EEG signal classification method according to claim 6, characterized in that: The specific operations of step S4 include: S401. The kernel feature matrix is ​​centralized using the centralization formula, which is: In the formula, Kx Take separately KS, KG, KGr(q) ; KS is the SPD manifold kernel characteristic matrix; KG is the Gaussian SPD manifold kernel feature matrix; KGr(q) The dimension of the linear subspace is q The Grassmann manifold kernel characteristic matrix when ; N is the number of segments of the MI-EEG signal; is a matrix with dimension N×N and all element values ​​are 1 / N; After centralization, the kernel feature matrices after centralization are obtained: ; S402. Perform dimensionality reduction processing on the core feature matrix after centralization, that is, perform eigenvalue decomposition, extract the eigenvectors corresponding to the first m eigenvalues, and retain the real part of the eigenvector to obtain three dimensionality reduction core feature matrices , the dimension of each dimensionality reduction kernel feature matrix is ​​N×j, each row of the dimensionality reduction kernel feature matrix represents a segment of MI-EEG signal, and each column represents a feature; S403. Perform L2 norm normalization on the reduced-dimensional kernel feature matrix according to the row feature vector. Suppose a row feature vector is , the corresponding L2 norm normalization formula is: Each row of feature vectors in the dimension reduction kernel feature matrix is ​​normalized according to the L2 norm normalization formula to obtain the normalized kernel feature matrix ; S404. Normalize the kernel feature matrix The concatenation and fusion are performed row by row to obtain a fusion kernel feature matrix with a dimension of N×3k. The expression of the fusion kernel feature matrix is: 。 8. A motor imagery EEG signal classification system based on multi-Riemann kernel fusion features, characterized in that: A method for classifying motor imagery EEG signals according to any one of claims 1 to 7, comprising: Model building and training module, used to build and train multi-Riemann kernel fusion feature classification model; A signal preprocessing module, used for preprocessing the MI-EEG signal to be classified to obtain a corresponding optimized MI-EEG signal; Multi-Riemann kernel feature extraction module, used to extract multi-Riemann kernel features from each segment of optimized MI-EEG signal, and obtain the kernel feature matrix of the optimized MI-EEG signal on SPD manifold, Gaussian SPD manifold and Grassmann manifold; The multi-Riemann kernel feature fusion module is used to perform centering, dimension reduction, and normalization processing on the kernel feature matrix of each optimized MI-EEG signal on the SPD manifold, Gaussian SPD manifold, and Grassmann manifold in turn, and perform feature fusion on the three processed kernel feature matrices to obtain a fused kernel feature matrix; The feature classification module is used to transmit the fusion kernel feature matrix to the multi-Riemann kernel fusion feature classification model for category prediction to obtain the category label of the MI-EEG signal.

9. An electronic device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor is used to implement the steps of the motor imagery EEG signal classification method as described in any one of claims 1 to 7 when executing the computer program.

10. A storage medium, characterized in that: A computer program is stored on the storage medium, and when the computer program is executed by the processor, the steps of the motor imagery EEG signal classification method as described in any one of claims 1 to 7 are implemented.

Images