A motor imagery electroencephalogram signal feature extraction method combining low rank representation and manifold learning

Abstract

This invention discloses a method for extracting features from motor imagery EEG signals by combining low-rank representation and manifold learning, comprising the following steps: S1: sampling motor imagery EEG signals to establish a sample set; S2: preprocessing the data using a filter bank co-space mode to obtain the original feature set of the data samples, and simultaneously dividing it into the original feature set of training samples and the original feature set of test samples; S3: further extracting data features using the proposed feature extraction method and performing dimensionality reduction; S4: classifying the extracted features using a support vector machine. The low-rank representation method effectively removes noise by separating the clean and noisy parts of the original data, and exhibits strong robustness to noise. Integrating low-rank representation, discriminative projection, and manifold learning into a unified model preserves both the global structural information and the local neighborhood relationships of the original samples, thereby acquiring more effective features and improving subsequent classification accuracy.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of motor imagery electroencephalogram signal classification and recognition, and particularly relates to a motor imagery electroencephalogram signal feature extraction method combining low-rank representation and manifold learning. BACKGROUND

[0002] The brain-computer interface records and analyzes the brain activity through a signal acquisition system, so as to realize the connection between the computer and the brain, and is currently mainly used for the physical rehabilitation of patients and the assistance of behavioral activities. The key to the application lies in the feature selection and extraction of electroencephalogram signals and the construction of a classification model. Not all features extracted from electroencephalogram signals are related to classification. Too many features not only increase the dimension of the feature matrix, but also lead to a low classification success rate. The electroencephalogram signal has the characteristics of nonlinearity, instability and weak signal, and the signal collected by the electrode is not simply vectorized data, but a multi-channel matrix data that collects the time sequence sampling of each electrode channel. The sampling data of each electrode channel is synthesized into the entire electroencephalogram signal. The conventional feature extraction method often ignores the nonlinear characteristics and spatial structure information in the signal. Therefore, how to denoise and enhance the original electroencephalogram data sample and solve the problems of high computational complexity and low classification accuracy caused by high-dimensional data is crucial. Manifold learning is a simulation of the information processing mechanism of human neural perception system, processing and memory of external things, and is essentially a nonlinear dimension reduction method, that is, a low-dimensional feature space is used to represent the geometric distribution of high-dimensional nonlinear data, which can maintain the information integrity of data in low-dimensional data and extract key information to discover the inherent rules of data itself. The basic idea is to assume that the original sample is located on the manifold structure in the high-dimensional data space, to keep the neighborhood relationship between the data unchanged, to map the high-dimensional data sample to the low-dimensional space through nonlinear transformation, and to achieve the purpose of data dimension reduction or visualization. In recent years, manifold learning algorithms have achieved a large number of research results and have shown certain advantages in the research of electroencephalogram signals.

[0003] Traditional dimension reduction methods only obtain the global structure information of observation data. In the manifold learning method, the local neighborhood structure of the original sample is well preserved, while the global structure information of the observation sample is ignored. The manifold learning method usually uses Euclidean distance to preserve the local structure information, and this distance is very sensitive to errors and noise in the observation data. Some scholars believe that noise-free data has a low-rank structure feature, and the method based on low-rank representation can explore the low-rank structure of the data, remove redundant information while maintaining the global structure of the data, and can effectively remove noise and have strong robustness to noise.

[0004] For example, a kind of "motor imagination eeg signal feature extraction method" disclosed in Chinese patent literature, its publication number: CN113239778A, its application date: May 10, 2021, can balance between the calculation complexity and convergence speed, it is convenient to realize in single-chip microcomputer, can satisfy the demand of eeg signal real-time acquisition. With fast convergence speed, it is not easy to change the waveform shape, it can effectively remove physiological artifacts and circuit noise effect, but there are problems of not effectively removing noise, incomplete global structure information, incomplete effective feature extraction and low classification accuracy. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application provides a motor imagination eeg signal feature extraction method combining low rank representation and manifold learning, which solves the problems of not effectively removing noise, incomplete global structure information, incomplete effective feature extraction and low classification accuracy.

[0006] The technical solution of the present application is as follows: a motor imagination eeg signal feature extraction method combining low rank representation and manifold learning, comprising the following steps:

[0007] S1: motor imagination eeg signal sampling is carried out, and a sample set is established;

[0008] S2: the data is preprocessed using filter bank common spatial pattern to obtain a data sample original feature set, and at the same time, the data sample original feature set is divided into a training sample original feature set and a test sample original feature set, and the maximum iteration number T and the dimension d are set;

[0009] S3: the proposed feature extraction method is used to further extract data features, and dimensionality reduction is carried out;

[0010] S4: the extracted features are classified by support vector machine.

[0011] As a preferred, the preprocessing method is: a chebyshev II filter bank covering 4-40Hz is used to divide all eeg signals into multiple sub-frequency bands, each 4Hz band is divided, and a total of 9 sub-frequency bands are divided.

[0012] As a preferred, the 9 frequency band data obtained after processing is used for feature extraction by OVR-CSP to obtain a group of frequency-space feature data sets.

[0013] As a preferred, an alternating iterative algorithm is designed to solve the data features after dimensionality reduction: fixing variable A to update Z and E, and fixing Z to update A; wherein A is the required projection matrix, Z and E are the low rank projection matrix and error matrix of the low rank representation part.

[0014] As a preferred, the method for updating Z and E is: initializing A, and then solving Z and E by using augmented Lagrange algorithm according to LRR algorithm.

[0015]

[0016] Solve the function with LRR algorithm, taking XA as the data matrix for low-rank decomposition;

[0017] Optimize with ALM algorithm:

[0018]

[0019] Minimize the following augmented Lagrange function:

[0020]

[0021] Where μ>0 is the penalty coefficient, ‖·‖ F is the Frobenious norm of matrix, Y1, Y2 are Lagrange multipliers. By fixing other variables respectively, then updating the Lagrange multipliers, the variables Z, J, E can be minimized one by one.

[0022] Solve the optimization J by singular value threshold operator:

[0023]

[0024] Optimal Z * :

[0025] Z * = (I + A T X T XA) -1 [A T X T (XA-E) + J + (A T X T Y1-Y2) / μ]

[0026] Solve the optimal E by matrix completion singular value threshold algorithm * :

[0027]

[0028] Update the multipliers and parameters as follows:

[0029] Y1 = Y1 + μ(XA-XAZ-F)

[0030] Y2 = Y2 + μ(Z-J)

[0031] μ = min(ρμ, μ max )

[0032] Where ρ is a parameter, and ρ>0.

[0033] As preferred, the method for solving A is. Let Xm = XZ, respectively construct the inter-class weight matrix B and the z intra-class weight matrix W, the element expression is respectively:

[0034]

[0035] Wherein, M i is the sample mean matrix of the i-th class sample, n i is the sample quantity of the i-th class sample; t is a parameter, by adjusting t can modify the weight; according to the following formula respectively calculate the intra-class scatter matrix P w and the inter-class scatter matrix P b :

[0036]

[0037]

[0038] Based on the constraint term XA=XAZ+E, the objective function can be converted to:

[0039]

[0040] Simplify as follows:

[0041]

[0042] Convert to generalized eigenvalue problem, sort the characteristic matrix obtained according to the eigenvalue size, then the projection matrix A can be obtained.

[0043] Among the above-mentioned parameters, α is set to 1, and λ is set to 0.01.

[0044] As preferred, update the matrix A, Z and E simultaneously until the specified maximum iteration number T is reached.

[0045] As preferred, the extraction method of d-dimensional feature data: according to Y=XA, the projected feature matrix Y is calculated, and the first d data are selected as the d-dimensional feature data after dimension reduction.

[0046] As preferred, the d-dimensional feature data is classified by using the SVM classifier.

[0047] The beneficial effects of the present application are: by using the low rank representation method, the global subspace structure of the original sample is captured, the lowest rank representation of the original electroencephalogram data is obtained, the original data is divided into clean part and noise part, the noise is effectively removed, and the noise has strong robustness; the low rank representation, discriminant projection and manifold learning method are integrated into a unified model, the global structure information of the original sample is retained, the local neighborhood relationship of the original sample is retained, the effective features are obtained to a greater extent, and the subsequent classification accuracy is improved. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 A logic flow diagram of the motor imagery electroencephalogram signal feature extraction method combining low rank representation and manifold learning of the present application. DETAILED DESCRIPTION

[0049] The technical solutions of the present application will be further specifically described below by examples and in combination with the drawings. In addition, in order to better illustrate the present application, a large number of specific details are given in the following. Those skilled in the art should understand that the present application can also be implemented without certain specific details. In some examples, methods and means familiar to those skilled in the art are not described in detail in order to highlight the main idea of the present application.

[0050] Example: In order to prove the feasibility of the above method, the algorithm is tested using the 2008 motor imagery data set Dataset 2a. The data set Data set 2a is composed of electroencephalogram data of nine healthy subjects, and the data extraction method is as follows: at the beginning of the experiment, a cross will appear on the black screen, prompting the subject to be ready, then an arrow in a certain direction will appear as a visual prompt for a few seconds, at this time the subject is required to perform a specific motor imagery task according to the prompt. After the visual prompt disappears, a short rest is performed, and the next experiment is waited for. Each subject is instructed to perform four different motor imagery tasks, i.e. left hand, right hand, both feet and tongue movement. Under a sampling frequency of 250HZ, 22 channels of EEG signals and 3 channels of EOG signals are recorded. The training data and test data of the subjects are extracted in two days.

[0051] In this embodiment, the specific steps are as follows:

[0052] Step 1: Perform motor imagery electroencephalogram signal sampling to establish a sample set.

[0053] Step 2: Preprocess the data using filter bank common spatial pattern to obtain the original feature set of the data sample Wherein The feature matrix X of each sample i has a size of N f Xg , total sample number is N, the matrix set corresponding to the category n s is the sample number in the s-th category, category s = 1, 2,..., z. At the same time, the training sample set and the test sample set are divided is the original feature set of the training sample and the original feature set of the test sample.

[0054] Firstly, all the electroencephalogram signals are divided into multiple sub-frequency bands by using a Chebyshev II filter bank covering 4-40 Hz, and each 4 Hz band is divided into one segment, and a total of 9 sub-frequency bands are divided, which are 4 Hz-8 Hz, 8 Hz-12 Hz, 12 Hz-16 Hz, 16 Hz-20 Hz, 20 Hz-24 Hz, 24 Hz-28 Hz, 28 Hz-32 Hz, 32 Hz-36 Hz and 36 Hz-40 Hz.

[0055] Then, the 9 frequency band data obtained after processing are subjected to one-versus-rest common spatial pattern algorithm (OVR-CSP) for feature extraction in the first stage, and a set of frequency-space feature data sets are obtained. OVR-CSP is a commonly used method for extending CSP to multi-class classification, which takes one of the modes as a class, and all the remaining modes as another class, thereby forming a new two-class, and then calculating the corresponding CSP for each class of mode. The z groups of features are combined, z is the total number of categories, and finally a feature matrix X with a size of N f ×N g is obtained, where N f is the number of divided frequency bands, N g = 2 * p * z, p is the feature vector selection number parameter in the calculation of the projection matrix in the CSP algorithm. Here, p is an integer, and the value range is set to 1-4, and the maximum iteration number T and the dimension d are set.

[0056] Step 3: Further extract data features using the proposed feature extraction method, and perform dimension reduction. The merit function of the proposed LR2DDLPP feature extraction method is:

[0057]

[0058] Where A is the projection matrix, and the projected feature is Y = XA.

[0059] In step 3, multiple variables are involved, and the optimal solution cannot be obtained at the same time, and an alternating iteration algorithm is designed to solve it. Update Z and E by fixing A, and update A by fixing Z. Where A is the required projection matrix, Z and E are the low-rank projection matrix and error matrix of the low-rank representation part.

[0060] First, fix A to solve Z and E. Initialize A, then solve Z and E by using the augmented Lagrangian method according to the LRR algorithm. At this time, the objective function is as follows:

[0061]

[0062] It can be found that it is a modified LRR problem, so the LRR algorithm is used to solve the function, and XA is regarded as a data matrix for low-rank decomposition. The optimization problem can be solved by using the ALM algorithm.

[0063] The above formula is converted into the following formula:

[0064]

[0065] The above problem can be solved by using the ALM method, and the objective is to minimize the following augmented Lagrangian function:

[0066]

[0067] In the above formula, μ>0 is a penalty coefficient, ‖·‖ F is the Frobenious norm of the matrix, Y1 and Y2 are Lagrange multipliers. By fixing other variables respectively, then updating the Lagrange multipliers, the variables Z, J and E can be minimized one by one.

[0068] J is optimized by the following formula:

[0069]

[0070] The above formula can be solved by the singular value threshold operator.

[0071] The optimal Z * can be updated by the following formula:

[0072] Z * = (I + A T X T XA) -1 [A T X T (XA-E) + J + (A T X T Y1-Y2) / μ]

[0073] The optimal E * can be updated by the following formula:

[0074]

[0075] The above formula can be solved by the singular value threshold algorithm of matrix completion.

[0076] The multipliers and parameters are updated as follows:

[0077] Y1 = Y1 + μ (XA - XAZ)

[0078] Y2 = Y2 + μ (Z - J)

[0079] μ = min (ρμ, μ max )

[0080] where p is a parameter and p > 0.

[0081] Then, fix Z to find p. When Z and E are fixed, let X m = XZ, construct the between-class weight matrix B and the z within-class weight matrices W, respectively, and the element expressions are as follows:

[0082]

[0083] where M i is the sample mean matrix of the i-th class sample, n i is the sample size of the i-th class sample. t is a parameter, and the weight can be modified by adjusting t.

[0084] Then, calculate the within-class scatter matrix P w and the between-class scatter matrix P b :

[0085]

[0086]

[0087] Based on the constraint term XA = XAZ + E, the objective function can be converted to:

[0088]

[0089] which can be further simplified as follows:

[0090]

[0091] Convert to a generalized eigenvalue problem, and sort the obtained characteristic matrix according to the eigenvalue size to obtain the projection matrix A.

[0092] Among the above-mentioned parameters, α is set to 1, and λ is set to 0.01.

[0093] Then, update the matrices A, Z and E simultaneously until the specified maximum number of iterations T is reached.

[0094] Then, calculate the projected feature matrix Y according to Y = XA, form a vector and select the first d data as the d-dimensional feature data after dimensionality reduction.

[0095] Step 4: Classification of the extracted features using support vector machines.

[0096] In this embodiment, only the EEG data of left and right hand motor imagery in the training stage is used in the experiment, which includes complete label information. The experimental process is set as follows:

[0097] The data set is preprocessed according to step S2 to obtain the original sample feature data set, and then the following three feature extraction methods are performed as comparative experiments, which are: no feature extraction processing, 2DDLPP method feature extraction and feature extraction of the method of the application, finally, the d-dimensional feature data is taken as the input of the SVM classifier for classification. The parameters d and p are both taken as the optimal values. The experimental results are shown in Table 1:

[0098] Table 1: Classification accuracy statistics table of comparative experiments.

[0099]

[0100] From the results of the comparative experiments, it can be seen that the method of the application shows the best classification accuracy among all the comparative algorithms, and the application provides a better method for feature extraction of motor imagery EEG signals, which preserves the global structure information and the local neighborhood relationship at the same time, and makes it have better robustness, and greatly improves the acquisition of discriminative features.

[0101] The above is only a specific embodiment of the application, but the protection scope of the application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the application, which should be covered within the protection scope of the application. Therefore, the protection scope of the application should be subject to the protection scope of the claims.

Claims

1. A motor imagery electroencephalogram signal feature extraction method combining low-rank representation and manifold learning, characterized in that, The method comprises the following steps: S1: motor imagery electroencephalogram signal sampling is performed, and a sample set is established; S2: the data is preprocessed using a filter bank common spatial pattern to obtain a data sample original feature set, and the data sample original feature set is divided into a training sample original feature set and a test sample original feature set; a maximum iteration number T and a dimension d are set; a one-versus-all common spatial pattern algorithm is used for feature extraction in a first stage; S3: a motor imagery electroencephalogram signal feature extraction method combining low-rank representation and manifold learning is used to further extract data features and perform dimension reduction; first, a variable A is fixed to update Z and E, and then Z is fixed to update A, wherein A is a required projection matrix, Z and E are low-rank projection matrices and error matrices of the low-rank representation part, and features are calculated according to the projection matrix A; S4: the extracted features are classified using a support vector machine. 2.The motor imagery electroencephalogram signal feature extraction method combining low rank representation and manifold learning according to claim 1, characterized in that, The preprocessing method is to divide all electroencephalogram signals into a plurality of sub-frequency bands, and the intervals of each sub-frequency band are equal. 3.The motor imagery electroencephalogram signal feature extraction method combining low-rank representation and manifold learning according to claim 2, characterized in that, Feature extraction is performed on the data of the plurality of sub-frequency bands to obtain a group of frequency-space feature data sets. 4.The motor imagery electroencephalogram signal feature extraction method combining low rank representation and manifold learning according to claim 1, characterized in that, The method for updating Z and E is: initializing A, and then solving Z and E by using the augmented Lagrange algorithm according to the LRR algorithm, The function is solved by using an LRR algorithm, and XA is regarded as a data matrix for low-rank decomposition; An ALM algorithm is used for optimization: An augmented Lagrange function is minimized as follows: where μ > 0 is a penalty coefficient, || · || is the Frobenious norm of a matrix, F is the Frobenious norm of a matrix, Y1, Y2 are Lagrange multipliers; by fixing other variables respectively, updating the Lagrange multipliers, minimizing the variables Z, J, E one by one; The optimization J is solved by a singular value threshold operator: Optimal Z * : Z * = (I + A T X T XA) -1 [A T X T (XA-E) + J + (A T X T Y1-Y2) / μ] Optimal E by matrix completion and singular value thresholding algorithm * : The multipliers and parameters are updated as follows: Y1 = Y1 + μ(XA-XAZ-E) Y2 = Y2 + μ(Z-J) μ = min(pμ, μ max ) Wherein ρ is a parameter, and ρ > 0.

5. The method of claim 4, wherein the method further comprises: The method for obtaining A is: let X m = XZ, respectively construct the inter-class weight matrix B and the z intra-class weight matrices W, and the element expressions are respectively: where M i is the sample mean matrix of the ith class, n i is the number of samples in the ith class; t is a parameter, which can modify the weight by adjusting t; the within-class scatter matrix P w and the between-class scatter matrix P b are calculated according to the following equations, respectively: Based on the constraint term XA = XAZ + E, the objective function can be converted to: Simplified as follows: The projection matrix A can be obtained after the obtained feature matrix is sorted according to the eigenvalue size; among the above-mentioned parameters, the value of α is set to 1, and the value of λ is set to 0.

01.

6. The method of claim 5, wherein the method further comprises: The matrices A, Z and E are updated at the same time until the specified maximum iteration number T is reached.

7. The method of claim 6, wherein the method further comprises: The d-dimensional feature data extraction method: the projected feature matrix Y is calculated according to Y = XA, the vector is drawn, and the first d data are selected as the d-dimensional feature data after dimension reduction. 8.The motor imagery electroencephalogram signal feature extraction method combining low-rank representation and manifold learning according to claim 7, characterized in that, The d-dimensional feature data is classified by using an SVM classifier.

Images