Motor imagery electroencephalogram signal feature extraction method and system

Abstract

The invention provides a motor imagery electroencephalogram signal feature extraction method and system, and relates to the technical field of electroencephalogram signal recognizing.The method comprises the following steps that original electroencephalogram data are obtained through a plurality of collection channels, and the original electroencephalogram data are preprocessed to obtain an electroencephalogram signal; performing various signal decomposition and FBCSP feature extraction on the preprocessed original electroencephalogram data to obtain a plurality of corresponding electroencephalogram signal features; splicing and combining the multiple electroencephalogram signal features in sequence, and screening by using mutual information to complete motor imagery electroencephalogram signal feature extraction; according to the method, the number of effective features is expanded, the signal-to-noise ratio can be improved, effective extraction of electroencephalogram information in a motor imagery related frequency band is ensured, and the classification accuracy of left and right hand motor imagery electroencephalogram signals is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of electroencephalogram (EEG) signal recognition, and more particularly to a method and system for extracting features of motor imagery EEG signals. Background Art

[0002] At present, brain-computer interface (BCI) provides a way for the human brain to control external devices, and provides a communication and control channel for people with severe movement disorders. It can convert the brain's neurophysiological signals into commands to control external devices. The motor imagery EEG signal refers to the signal recorded by the electroencephalogram (EEG) when an individual is performing motor imagery. Research in this field focuses on how to identify the individual's motor imagery state through EEG signals, and this research is of great significance in brain-computer interface (BCI) and cognitive neuroscience.

[0003] Research on motor imagery has shown that during unilateral arm movement or imagined movement, there will be attenuation or enhancement of power spectral density in specific frequency bands, such as alpha waves (8-13Hz) and beta waves (14-30Hz), which is called event-related desynchronization (ERD) and event-related synchronization (ERS). Based on this phenomenon, researchers have designed various algorithms to extract feature information from different types of motor imagery EEG signals to facilitate subsequent identification and classification. Among them, the common spatial pattern (CSP) is a commonly used spatial filter suitable for binary classification. It constructs the optimal spatial filter to maximize the variance difference of the multi-channel EEG signal matrix and obtains a discriminative feature vector from it.

[0004] However, the feature extraction of EEG signals for left and right hand movement imagination in the existing technology faces many challenges, such as the non-stationary characteristics of EEG signals, low signal-to-noise ratio (SNR), and artifacts such as electrooculogram (EOG), electromyogram (EMG) and power line interference. The combined effect of these factors makes it quite difficult to accurately identify movement intentions from recorded EEG signals. The disadvantage of the CSP method is that it has a weak resistance to noise interference. How to suppress noise is an important way to improve the accuracy of the CSP algorithm.

[0005] Therefore, how to provide a method for extracting motor imagery EEG signal features that can solve the above problems is an urgent problem that those skilled in the art need to solve. Summary of the invention

[0006] In view of this, the present invention provides a method and system for extracting features of motor imagery EEG signals, which extract features of motor imagery EEG signals based on two different signal decomposition methods, namely intrinsic time-scale decomposition (ITD) and CEEMDAN decomposition, in combination with filter bank common spatial pattern (FBCSP), namely the ITD+CEEMDAN-FBCSP method, which expands the number of effective features, can improve the signal-to-noise ratio, ensure the effective extraction of EEG information in the frequency band related to motor imagery, and improve the classification accuracy of left and right hand motor imagery EEG signals.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] A method for extracting features of motor imagery EEG signals, comprising the following steps:

[0009] Acquire raw EEG data through multiple acquisition channels, preprocess the raw EEG data, and simultaneously obtain the correlation coefficient between the amplitude of the preprocessed raw EEG signal of each acquisition channel and the label value, and select the raw EEG data of the required channel by comparing the correlation coefficient with a preset threshold value;

[0010] Performing multiple signal decompositions and FBCSP feature extraction on the preprocessed raw EEG data to obtain corresponding multiple EEG signal features;

[0011] The plurality of EEG signal features are sequentially spliced ​​and combined, and are screened using mutual information to complete the extraction of motor imagery EEG signal features.

[0012] Preferably, the specific processing process of obtaining the corresponding multiple EEG signal features includes:

[0013] Perform ITD decomposition and CEEMDAN decomposition on the preprocessed raw EEG data to obtain corresponding ITD decomposition results and CEEMDAN decomposition results;

[0014] Selecting the i-layer component of the ITD decomposition result and the i-layer component of the CEEMDAN decomposition result, where i=1, 2, 3, 4, to obtain the corresponding ITD component and CEEMDAN component;

[0015] Performing narrowband filtering on the ITD component and the CEEMDAN component, wherein the ITD component and the CEEMDAN component after the narrowband filtering have the same dimension as the original EEG data;

[0016] The spatial filters corresponding to the ITD decomposition results and the CEEMDAN decomposition results after narrowband filtering are respectively constructed, and the corresponding spatial filters are used to extract features of the ITD components and CEEMDAN components after narrowband filtering to obtain corresponding multiple EEG signal features.

[0017] Preferably, it also includes:

[0018] Using mutual information to screen the multiple EEG signal features that have been spliced ​​and combined to obtain a screening result;

[0019] A classifier model is constructed, and the screening results are input into the classifier model for classification to obtain corresponding classification results.

[0020] The present invention also provides a system for extracting features of motor imagery EEG signals, comprising:

[0021] An acquisition module is used to acquire raw EEG data through multiple acquisition channels, preprocess the raw EEG data, and simultaneously acquire the correlation coefficient between the amplitude of the preprocessed raw EEG signal of each acquisition channel and the label value, and select the raw EEG data of the required channel by comparing the correlation coefficient with a preset threshold value;

[0022] A feature extraction module, used for performing multiple signal decomposition and FBCSP feature extraction on the pre-processed raw EEG data to obtain corresponding multiple EEG signal features;

[0023] A feature combination and feature selection module is used to sequentially combine the multiple EEG signal features and screen them using mutual information to complete the extraction of motor imagery EEG signal features;

[0024] The classification module is used to build a classifier model and output the predicted classification results.

[0025] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a method and system for extracting features of motor imagery EEG signals, which has the following beneficial effects:

[0026] (1) In view of the previous situation of using only EMD signal decomposition or only ITD signal decomposition, the present invention integrates two signal decomposition strategies, ITD and CEEMDAN decomposition, to further expand the number of effective features. This solves the problem of low classification accuracy of traditional CSP when there are fewer leads, while retaining the characteristic of further improving the classification accuracy as the number of leads collected in the motor imagery brain area increases.

[0027] (2) In view of the shortcomings of the traditional CSP feature extraction algorithm, the filter bank common spatial pattern (FBCSP) algorithm adopted in the present invention adds a filter bank on the original basis, divides the signal into multiple sub-band signals, and then uses the CSP algorithm to extract the spatial energy characteristics of each filtered sub-band signal, thereby realizing the effective extraction of EEG information within a smaller target frequency scale range, increasing the signal-to-noise ratio, and realizing individual adaptation of the filter to a certain extent.

[0028] (3) In view of the limitation that the subjects need to have motor imagery training experience in the traditional left-hand and right-hand EEG signal recognition process, the present invention also shows good classification performance for subjects without such foundation, laying a solid foundation for the practical application of portable BCI systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0030] Figure 1 A flowchart of a method for extracting features of motor imagery EEG signals provided by an embodiment of the present invention;

[0031] FIG. 2( a ) and FIG. 2( b ) are spectrograms of the first four components of an EEG signal of any lead after ITD decomposition and CEEMDAN decomposition, respectively, provided in an embodiment of the present invention;

[0032] FIG. 3( a ) and FIG. 3( b ) are schematic diagrams of the experimental paradigm and lead distribution of the 128-channels dataset of untrained subjects collected in a laboratory provided by an embodiment of the present invention;

[0033] Figure 4 A structural principle block diagram of a motor imagery EEG signal feature extraction system provided by the present invention. DETAILED DESCRIPTION

[0034] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0035] See also Figure 1As shown, an embodiment of the present invention provides a method for extracting features of motor imagery EEG signals, comprising the following steps:

[0036] The original EEG data is obtained through multiple acquisition channels, wherein the correlation coefficient between the amplitude of the pre-processed original EEG signal of each acquisition channel and the label value is obtained, and the original EEG data of the required channel is selected by comparing the correlation coefficient with the preset threshold value. By reducing the number of irrelevant channels, the data dimension to be processed in the subsequent process is reduced, and the channels with strong correlation are retained. The EEG signals generated by the brain regions corresponding to these channels store effective information that is highly correlated with the motor imagery task;

[0037] The raw EEG data is preprocessed, wherein the preprocessing process can use a finite impulse response (FIR) filter to perform a 8-30 Hz bandpass filter on the raw EEG data. The purpose of selecting the 8-30 Hz bandpass filter is to retain more effective components from the raw EEG data. At the same time, in the process of obtaining the raw EEG data, the number of collected leads in the sensorimotor area and the supplementary motor area can be increased to ensure higher collection accuracy and classification effect. The lead data collection process can be shown in Figures 3(a) and 3(b);

[0038] The preprocessed EEG data are subjected to multiple signal decompositions and FBCSP feature extraction to obtain corresponding multiple EEG signal features;

[0039] Multiple EEG signal features are spliced ​​and combined in sequence, and mutual information is used for screening to complete the extraction of motor imagery EEG signal features.

[0040] In a specific embodiment, the specific processing process of obtaining the corresponding multiple EEG signal features includes:

[0041] Perform ITD decomposition and CEEMDAN decomposition on the preprocessed EEG data to obtain the corresponding ITD decomposition results and CEEMDAN decomposition results;

[0042] Select the i-layer component (PRCi) of the ITD decomposition result and the i-layer component (IMFi) of the CEEMDAN decomposition result, where i = 1, 2, 3, 4, and obtain the corresponding ITD component and CEEMDAN component. The specific results are shown in Figure 2(a) and Figure 2(b);

[0043] The ITD component and the CEEMDAN component are subjected to narrowband filtering, wherein the frequency band of the filter group can be set to [8-10.5] Hz, [10.5-15] Hz, [15-18] Hz, and [18-25] Hz during the narrowband filtering process, and the ITD component and the CEEMDAN component subjected to narrowband filtering have the same dimension as the original EEG data;

[0044] The spatial filters corresponding to the ITD decomposition results and the CEEMDAN decomposition results after narrowband filtering are constructed respectively, and the corresponding spatial filters are used to extract features of the ITD components and CEEMDAN components processed by narrowband filtering to obtain corresponding multiple EEG signal features.

[0045] Specifically, the embodiment of the present invention applies the ITD decomposition algorithm and the CEEMDAN decomposition algorithm to decompose the filtered EEG signal, and retains the first 4 layers of components obtained after ITD decomposition and the first 4 layers of components obtained after CEEMDAN decomposition, so as to fully include the relevant information of the motor imagery frequency band.

[0046] Specifically, ITD decomposition, namely intrinsic time-scale decomposition (ITD), is suitable for analyzing non-stationary and nonlinear signals. It does not rely on spline interpolation to the signal extreme points and the screening process in single component separation, but decomposes the signal into proper rotation components (PRCs), which are based on the baseline defined by linear transformation and are suitable for calculating instantaneous frequency and amplitude. The decomposition result obtained accurately retains the time information of each component about the critical points and wave riding phenomenon of the signal, and its time resolution is equal to the time scale of the extreme points in the original signal. It is based on the analysis of a single wave and can accurately extract the inherent instantaneous amplitude, frequency / phase information and other related morphological features. The specific process includes:

[0047] For the preprocessed EEG signal I(t), the operator L is defined to extract the baseline signal from the EEG signal I(t), and the remaining signal is called the Proper Rotation Component (PRC). The decomposition of the EEG signal I(t) can be expressed as:

[0048] I(t)=LI(t)+(1-LI(t))=B(t)+H(t)

[0049] Where B(t) represents the baseline signal and H(t) is the appropriate rotation component.

[0050] The decomposition process of nonlinear signals can be summarized as the following steps:

[0051] (1) Find the local extreme point of the signal I(t), denoted as I k , and the corresponding appearance time instant τ k , k=0,1,2,...., for convenience, let τ0=0;

[0052] (2) Assume that in the interval [0,τ k ], and the signal I(t) is set in the interval [0,τ k+2 ], then at two adjacent extreme points I k and I k+2 The interval between [τ k ,τ k+1 ], the segmented baseline extraction operator is defined as:

[0053]

[0054] In the formula And 0<β<1, β is usually taken as 0.5.

[0055] (3) After extracting the baseline signal, an operator θ is defined to extract the remaining signal as PRCs, which is defined as:

[0056] ΘI(t)≡(1-L)I(t)=I(t)-B(t);

[0057] By definition, PRC is the wave with the highest frequency on the baseline. Therefore, ITD separates PRCs in order of frequency from high to low. Take the baseline B(t) as the input signal I(t) and repeat steps (1) to (3) until the baseline becomes a monotonic function or a constant. Finally, the original signal will be decomposed into a series of PRCs and a trend. The specific expression is:

[0058]

[0059] Here, ρ represents the number of decomposition levels.

[0060] In an embodiment of the present invention, an intrinsic time scale decomposition (ITD) algorithm is applied to the filtered electroencephalogram (EEG) signal obtained from the selected lead. After retaining the first four properly rotated components (PRCs), each of the four PRCs is individually processed by a filter bank common spatial patterns (FBCSP) algorithm for feature extraction.

[0061] Specifically, CEEMDAN decomposition is to add adaptive white noise to each stage of the integrated intrinsic mode decomposition EEMD technology, so that the decomposition process is complete, while effectively suppressing the noise components in the motor imagery EEG signal to achieve the purpose of high-quality denoising. This method can effectively solve the frequency confusion problem in the signal decomposition process of the traditional intrinsic mode decomposition EMD technology, and also improve the problem of residual white noise of the intrinsic mode component in the ensemble intrinsic mode decomposition EEMD technology. The uniqueness of CEEM-DAN is that it introduces a pair of positive and negative Gaussian white noise into the signal to be processed, and then uses the intrinsic mode decomposition EMD technology to decompose the signal. The detailed process can be seen in the following formula:

[0062]

[0063] Where z(t) represents the original signal, represents the standard deviation of the noise, vi represents Gaussian white noise following the standard distribution, q represents the introduced bidirectional noise, C1(t) represents the intrinsic mode function component in the initial stage, denoted as IMF1, and r1(t) represents the remaining signal part.

[0064] In order to calculate the mean and residual of each intrinsic mode function component, the study refers to the following formula.

[0065]

[0066] After adding a pair of Gaussian white noise to the residual signal again, an updated signal is formed. This signal is then decomposed through the intrinsic mode decomposition (EMD) process to obtain the second-layer internal mode function component C2(t), denoted as IMF2. Continuing this process, the mean and residual of the third-layer internal mode function component are further extracted as shown in the following formula.

[0067]

[0068] Continue the above process until the remaining residual signal becomes a monotonic function that cannot be further decomposed, which marks the completion of the CEEMDAN algorithm. At this stage, the original signal z(t) is completely decomposed into the following formula:

[0069]

[0070] Where z(t) is the original reconstructed signal, C1(t), C2(t), ..., C p (t) is the IMF component decomposed by CEEMDAN, p is the number of specified IMFs, r i (t) is the residual signal of each layer.

[0071] In an embodiment of the present invention, a complete set of intrinsic mode decomposition (CEEMDAN) algorithm is applied to the filtered electroencephalogram (EEG) signal obtained from the selected lead. After retaining the first four intrinsic modes (IMFs), each of the four IMFs is individually processed by a filter bank common spatial patterns (FBCSP) algorithm for feature extraction.

[0072] In the embodiment of the present invention, the 8 component signals obtained by decomposition are filtered on 4 sub-bands of [8-10.5Hz], [10.5-15Hz], [15-18Hz], and [18-25Hz], and then CSP feature extraction is performed, wherein each component can obtain 2m features in each sub-band. In the embodiment of the present invention, the value of m is 3. These features are spliced ​​to obtain a total of 192 features (8*4*6).

[0073] Specifically, the filter bank common spatial pattern (FBCSP) is an extended algorithm based on the common spatial pattern (CSP). First, a filter bank with multiple sub-band filters is used to divide the results from ITD or CEEMDAN decomposition into sub-bands in different frequency bands. Then, CSP (common spatial pattern) feature extraction is performed on each sub-band. The CSP method is a spatial domain filtering feature extraction algorithm under two-category classification tasks. This algorithm can extract spatial distribution components from multi-channel EEG (electroencephalogram) data. The principle of the CSP method is to find a set of optimal spatial filters by diagonalizing the matrix, maximize the difference in the variance of the two types of signals by projection, and then obtain a highly discriminative feature vector. The specific steps are summarized as follows:

[0074] ① Calculate the covariance matrix R L and R R as follows:

[0075]

[0076] In the formula, X L and X R are the EEG signals of left-hand motor imagery and right-hand motor imagery, respectively, tr(.) represents the trace of the matrix;

[0077] ② Calculate the covariance matrix R of the mixed space and perform eigenvalue decomposition:

[0078]

[0079] In the formula, and denote the average covariance matrices of the left and right hand classification tasks, respectively, U is the eigenvector matrix of R, and λ is the diagonal matrix of the corresponding eigenvalues.

[0080] ③Calculate the whitening matrix:

[0081]

[0082] ④ For R L and R R Perform whitening and principal component analysis (PCA):

[0083]

[0084] In the formula, S L and S R The eigenvectors of are the same, so when the eigenvalue of one matrix is ​​the largest, the eigenvalue of the other matrix will also be the smallest, which helps with classification.

[0085] ⑤Construct spatial filter W:

[0086] W=B T P

[0087] The embodiment of the present invention filters the PRC component or IMF component in each sub-band through a spatial filter W to obtain a characteristic matrix. Then, the first m and last m eigenvalues ​​are selected from the matrix to form a signal (where m=3), and the corresponding eigenvector is obtained:

[0088]

[0089] In a specific embodiment, it also includes:

[0090] After the above motor imagery EEG signal features are combined, they are screened using mutual information, that is, the mutual information between each feature and the label is calculated, and the k features with the largest mutual information value are screened from the combined features as the input of the classifier model;

[0091] Construct a classifier model, where the classifier model can be an SVM classifier, obtain a data set and divide it into a training set and a test set in proportion, use the training set to train the classifier model, use the test set to obtain the predicted classification result, verify the performance of the classifier model, and determine the optimal value of the parameter k according to the classification accuracy, that is, select the k value corresponding to the highest classification accuracy;

[0092] The screening results are input into the classifier model for classification to obtain the corresponding classification results.

[0093] In a specific embodiment, the specific process of building and improving the classification model includes:

[0094] The SVM classifier is optimized using the RBF radial basis kernel function and the error penalty factor.

[0095] Specifically, the embodiment of the present invention uses mutual information for feature selection: In information theory, mutual information is an extension of information entropy and is often used to measure the correlation between two random variables. The entropy of two time series and their joint entropy are defined as H(X), H(Y) and H(X,Y), respectively, and the corresponding calculation formula is as follows:

[0096]

[0097] Where p(x,y) is the joint probability density of X and Y, p x (y), p y (y) are the marginal probability densities of X and Y respectively, and the mutual information of X and Y is defined as:

[0098] I(X,Y)=H(X)+H(Y)-H(X,Y)

[0099] When X and Y are independent, I(X,Y)=0.

[0100] The embodiment of the present invention applies mutual information to CSP feature selection, where X and Y represent a CSP feature and a label respectively, so as to calculate the size of the mutual relationship between the feature and the label, so as to screen out some CSP features containing the most useful information and the most distinguishing.

[0101] Specifically, support vector machine (SVM) is a machine learning method based on statistical learning theory. It maps the input vector to a high-dimensional feature space through appropriate nonlinear mapping, so that the data (belonging to two categories) can always be separated by a hyperplane. The optimal classification surface constructed requires that the classification surface can not only correctly separate the two types of data, but also maximize the classification interval. For nonlinear problems, it is only necessary to transform it into a linear problem in another space through nonlinear transformation, and then construct the optimal classification surface.

[0102] During the training process of the SVM classifier, the features of a total of 112 trials can be divided into a training set and a test set according to the ratio, and the division ratio can be 9:1. At this time, the data dimension of the training set is K*100, and the dimension of the test set is K*12, where K is the number of features selected during feature selection; the K data of the training set and the corresponding labels are used to train the SVM model, and the K data of the test set are input into the trained SVM model to obtain the corresponding test results, which are then compared with the actual test results of the test set. The ratio of the number of correct predictions to the test results of the test set is statistically calculated as the accuracy (ACC), and a ten-fold cross-validation is performed on the above content, that is, a different test result is selected each time as the test set to calculate the corresponding accuracy, and the average result of the 10 accuracy rates is taken as the final accuracy under the corresponding K value to complete the classifier training.

[0103] The kernel parameter γ and the error penalty factor C are the main parameters that affect the performance of SVM; the value of γ affects the data distribution after spatial transformation, and the penalty factor C determines the convergence speed and generalization ability of the support vector machine. Therefore, the selection of γ and C greatly affects the recognition rate of EEG signals. The embodiment of the present invention uses the fitcsvm function in the MATLAB machine learning toolbox to construct a support vector machine model, using C=1.

[0104] See also Figure 4 As shown, the present invention also provides a motor imagery EEG signal feature extraction system, comprising:

[0105] An acquisition module is used to acquire raw EEG data through multiple acquisition channels, preprocess the raw EEG data, and simultaneously acquire the correlation coefficient between the amplitude of the preprocessed raw EEG signal of each acquisition channel and the label value, and select the raw EEG data of the required channel by comparing the correlation coefficient with a preset threshold value;

[0106] A feature extraction module, used for performing multiple signal decomposition and FBCSP feature extraction on the pre-processed raw EEG data to obtain corresponding multiple EEG signal features;

[0107] A feature combination and feature selection module is used to sequentially combine the multiple EEG signal features and screen them using mutual information to complete the extraction of motor imagery EEG signal features;

[0108] The classification module is used to build a classifier model and output the predicted classification results.

[0109] In order to verify the feasibility of the method provided by the embodiment of the present invention, an ablation experiment was conducted on the factor of "whether to perform ITD and CEEMDAN frequency band decomposition" on a self-collected 128-lead EEG data set (each subject was not trained). Taking the results under 40-lead signals as an example, as shown in Table 1, the results show that the classification results of frequency band decomposition are better than those of only performing FBCSP feature extraction without frequency band decomposition, and the average accuracy is improved by 22.5%.

[0110] Table 1 Verification results

[0111]

[0112] On a self-collected 128-lead EEG dataset (each subject was not trained), the classification accuracy of the algorithm ITD&CEEMDAN-FBCSP-SVM of the present invention was compared when different numbers of leads were selected. The results showed that when 40 leads were selected, the sensorimotor area was completely included, and the highest classification accuracy could be obtained, which reflects the effectiveness of the lead selection obtained by the method provided in the embodiment of the present invention. The specific results are shown in Table 2.

[0113] Table 2 Results of different lead selections

[0114]

[0115] On a self-collected 128-lead EEG data set (each subject was not trained), the classification accuracy of the algorithm ITD&CEEMDAN-FBCSP-SVM of the present invention was compared with that of the algorithm using only ITD decomposition or only CEEMDAN decomposition followed by FBCSP feature extraction and SVM classification (i.e., ITD-FBCSP-SVM or CEEMDAN-FBCSP-SVM). Taking the results under 40-lead signals as an example, the results show that compared with using only ITD decomposition or only CEEMDAN decomposition, the combination of ITD and CEEMDAN decomposition methods can obtain a higher classification accuracy, which reflects the effectiveness of the classification after combining the two signal decomposition methods of ITD and CEEMDAN in the algorithm of the present invention. The specific results are shown in Table 3.

[0116] Table 3 Verification results

[0117]

[0118] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0119] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one 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. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for extracting features of motor imagery EEG signals, characterized in that: The following steps are involved: Acquire raw EEG data through multiple acquisition channels, preprocess the raw EEG data, and simultaneously obtain the correlation coefficient between the amplitude of the preprocessed raw EEG signal of each acquisition channel and the label value, and select the raw EEG data of the required channel by comparing the correlation coefficient with a preset threshold value; Performing multiple signal decompositions and FBCSP feature extraction on the preprocessed raw EEG data to obtain corresponding multiple EEG signal features; The plurality of EEG signal features are sequentially combined and screened using mutual information to complete the extraction of motor imagery EEG signal features; The specific processing process of obtaining the corresponding multiple EEG signal features includes: Perform ITD decomposition and CEEMDAN decomposition on the preprocessed raw EEG data to obtain corresponding ITD decomposition results and CEEMDAN decomposition results; Selecting the i-th layer component of the ITD decomposition result and the i-th layer component of the CEEMDAN decomposition result, where i=1, 2, 3, 4, to obtain the corresponding ITD component and CEEMDAN component; Performing narrowband filtering on the ITD component and the CEEMDAN component, wherein the ITD component and the CEEMDAN component after the narrowband filtering have the same dimension as the original EEG data; The spatial filters corresponding to the ITD decomposition results and the CEEMDAN decomposition results after narrowband filtering are respectively constructed, and the corresponding spatial filters are used to extract CSP features from the ITD components and CEEMDAN components after narrowband filtering to obtain corresponding multiple EEG signal features.

2. The method for extracting motor imagery EEG signal features according to claim 1, characterized in that: Also includes: Using mutual information to screen the multiple EEG signal features that have been spliced ​​and combined to obtain a screening result; A classifier model is constructed, and the screening results are input into the classifier model for classification to obtain corresponding classification results.

3. A system using the method for extracting features of motor imagery EEG signals according to any one of claims 1 to 2, characterized in that: include: An acquisition module is used to acquire raw EEG data through multiple acquisition channels, preprocess the raw EEG data, and simultaneously acquire the correlation coefficient between the amplitude of the preprocessed raw EEG signal of each acquisition channel and the label value, and select the raw EEG data of the required channel by comparing the correlation coefficient with a preset threshold value; A feature extraction module, used for performing multiple signal decomposition and FBCSP feature extraction on the pre-processed raw EEG data to obtain corresponding multiple EEG signal features; A feature combination and feature selection module is used to sequentially combine the multiple EEG signal features and screen them using mutual information to complete the extraction of motor imagery EEG signal features; The classification module is used to build a classifier model and output the predicted classification results.

Images