Motor imagery electroencephalogram decoding method based on adaptive frequency band selection

Through adaptive band selection and feature extraction algorithms, the frequency band selection problem caused by individual differences in the brain-computer interface is solved, the accuracy and efficiency of EEG signal decoding are improved, and the individualized and efficient decoding effect is achieved.

CN119066500AInactive Publication Date: 2024-12-03ANYANG INST OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411052573.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2024-04-28
Filing Date
2024-08-02
Publication Date
2024-12-03
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

There are significant individual differences in the frequency domain of brain-computer interfaces based on motion imagination, resulting in the traditional fixed frequency band selection method affecting feature extraction and pattern classification effects, making it difficult to achieve high-precision electroencephalopathy.

Method used

Adaptive band selection method is adopted to calculate the power spectral density and peak value of the EEG signal, dynamically adjust the bandwidth, combine the co-spatial mode algorithm and optimal feature selection of mutual information, and use the support vector machine classifier for electroencephalogram decoding.

Benefits of technology

It improves the classification accuracy of brain-computer interfaces, realizes individualized efficient and accurate decoding, and adapts to the frequency band characteristics of different individuals.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119066500A_ABST
    Figure CN119066500A_ABST
Patent Text Reader

Abstract

The invention discloses a motor imagery electroencephalogram decoding method based on adaptive optimal frequency band selection, and belongs to the field of brain-computer interfaces (BCI). The method comprises the following steps: selecting key channels of motor imagery electroencephalogram signals; preprocessing the electroencephalogram data; selecting an adaptive optimal frequency band based on a power spectrum peak value; electroencephalogram signal feature extraction and feature selection; and recognizing the motor imagery intention to obtain a classification result. The method starts from the nature that the electroencephalogram signals induced by motor imagery present rhythmicity features, frequency band division is performed on the electroencephalogram signals in a self-adaptive mode, more remarkable BCI features are extracted, and the intention recognition effect can be effectively improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of brain-computer interface, and in particular to a motor imagery electroencephalogram decoding method based on adaptive frequency band selection. Background Art

[0002] Brain Computer Interface (BCI) is a new human-computer interaction mode that decodes the subjective intention of the brain into control instructions to achieve control with external devices. Among them, the brain-computer interface based on motor imagery is one of the important paradigms in the field of brain-computer interface. It can activate the neural oscillations of the human brain's motor perception cortex through motor imagery of different limb movements, thereby producing event-related desynchronization (ERD) and event-related synchronization (ERS) phenomena. Studies have shown that this neural oscillation has obvious rhythmicity, mainly concentrated in the frequency band of [8,36] Hz, among which the α band and β band are particularly obvious. At the same time, the brain-computer interface based on motor imagery helps to promote the functional remodeling of damaged cortical nerves, so it is also of great value in the field of neurorehabilitation.

[0003] However, EEG signals based on motor imagery have significant individual differences in the frequency domain, so high-precision brain-computer interface decoding depends on highly adaptive frequency band selection. At present, the commonly used frequency band selection is to use a multi-band filter bank to divide the EEG signal into multiple bands (such as FBCSP). However, the selected frequency bands and the number of frequency bands are artificially set. This fixed frequency band division strategy directly affects the subsequent feature extraction and pattern classification effects. Summary of the invention

[0004] The purpose of the present invention is to provide a motor imagery EEG decoding method based on adaptive frequency band selection based on the essence of rhythmic oscillation of motor imagery EEG signals, and to achieve efficient and accurate decoding of the brain-computer interface through adaptive selection of the frequency band of motor imagery EEG signals.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A motor imagery EEG decoding method based on adaptive frequency band selection, which selects the key channels of motor imagery EEG signals and preprocesses them; then calculates the power spectrum density of each channel signal, and uses the power spectrum peak to adaptively select the optimal frequency band; then uses the selected frequency band to perform multi-band decomposition of the EEG signal, uses the common space pattern algorithm (CSP) to extract spatial features from each sub-band data, and merges the features of each sub-band; uses the mutual information optimal feature selection algorithm (MIBIF) to select the merged feature set; finally, uses a support vector machine (SVM) classifier to perform EEG decoding. Specifically, it includes the following steps:

[0007] S1. Key channel selection: Select key channels for motor imagery EEG signals and preprocess the selected key channels;

[0008] S2, optimal frequency band selection: for the EEG data preprocessed in S1, the task-induced average power spectrum curve is calculated, and the optimal frequency band is adaptively selected according to the average power spectrum peak value;

[0009] S3, EEG signal feature extraction and feature selection: Use the optimal frequency band selected in S2 to divide the training set and test set data into frequency bands; extract the features of each sub-band, and select the merged sub-band features;

[0010] S4. Motor imagery feature classification: Use support vector machine (SVM) classifier to classify features to achieve motor imagery pattern recognition.

[0011] Preferably, the preprocessing in S1 specifically includes segmenting the EEG data, re-referencing using a common average reference (CAR), and removing power frequency interference and baseline drift.

[0012] Preferably, S2 specifically includes the following contents:

[0013] S2.1, construct a fifth-order Butterworth bandpass filter with a passband frequency of 8-36 Hz to perform bandpass filtering on the preprocessed EEG signal;

[0014] S2.2, use the Welch method to calculate the power spectral density of each channel signal;

[0015] S2.3. Average the power spectra of key channels in brain regions activated by different motor imagery tasks to obtain the task-induced average power spectrum curve;

[0016] S2.4, smoothing the average power spectrum curve to remove pseudo peaks caused by slight fluctuations in the curve;

[0017] S2.5, obtaining the power spectrum peak value of the average power spectrum curve, wherein the number of peak values ​​is the number of frequency bands to be divided;

[0018] S2.6. Dynamically adjust and determine the bandwidth of each power spectrum peak.

[0019] Preferably, S2.3 further includes the following contents: the brain area activated by left-hand motor imagery is located in the right brain area, and the brain area activated by right-hand motor imagery is located in the left brain area; the power spectrum curves of all key channels of the right brain area during the left-hand motor imagery task and the left brain area during the right-hand motor imagery task are averaged to obtain the task-induced average power spectrum curve.

[0020] Preferably, the S2.6 further includes the following contents:

[0021] S2.6.1. For each power spectrum peak point determined, the frequency value f corresponding to the peak point is 0 As the center, with H as the step length, determine the initial frequency band of the peak point [f 0 -H,f 0 +H]; the co-spatial pattern algorithm was used to extract EEG features within the initial frequency band, and the support vector machine (SVM) classifier was used for feature classification to obtain the classification accuracy acc 0 ;

[0022] S2.6.2, with H as the step size, increase the bandwidth to [f 0 -2*H,f 0 +2*H], and use the common spatial pattern algorithm (CSP) to extract features, and use the support vector machine (SVM) classifier to classify features to obtain the classification accuracy acc 1 .

[0023] S2.6.3, if acc 1 ≥acc 0 , stop increasing the frequency band, and the bandwidth used in S2.6.2 is the frequency band bandwidth of the peak point; otherwise, repeat S2.6.2 to increase the bandwidth, calculate the new classification accuracy, and update acc 1 ;

[0024] S2.6.4, when the lower limit frequency f 0 -n*H≤8Hz or upper limit frequency f 0 When +n*H≥36Hz, the frequency band is stopped from increasing; at the same time, f 0 -n*H is 8Hz or f 0 +n*H is 36Hz, calculate the classification accuracy and set it as acc 1 ; if acc 1 ≥acc 0, then the bandwidth at this time is the bandwidth of the peak point; otherwise, acc 0 The corresponding bandwidth is the peak point frequency band bandwidth.

[0025] Preferably, the frequency band step length H is determined by the number of peak points in the power spectrum. If the number of peak points is greater than 2, the step length H is 0.5 Hz; if the number of peak points is less than or equal to 2, the step length H is 1 Hz.

[0026] Preferably, S3 specifically includes the following contents:

[0027] S3.1, using the selected adaptive optimal frequency band, decompose the EEG signals of the training set and the test set into multiple sub-bands respectively;

[0028] S3.2, using the common spatial pattern algorithm (CSP) to extract features from each sub-band data;

[0029] S3.3, merging multiple sub-band features of the training set data and the test set data into a training set feature vector and a test set feature vector respectively;

[0030] S3.4. Use the mutual information optimal feature selection algorithm (MIBIF) to select the best features and obtain the best features of the training set and the best features of the test set.

[0031] Preferably, extracting the features of each sub-band in S3.2 specifically refers to extracting features of each sub-band data using a common spatial pattern algorithm (CSP), which specifically includes the following contents:

[0032] Assume X 1 and X 2 are two different types of data for motor imagery signals, and their covariance matrices are:

[0033]

[0034]

[0035] In the formula, Respectively represent X 1 and X 2 The transpose of , trace(·) represents the trace of the matrix;

[0036] Perform eigendecomposition on the sum of the two covariance matrices R:

[0037] R=R 1 +R 2 =UΛU T (3)

[0038] Among them, Λ is the eigenvalue diagonal matrix of matrix R, and U is the eigenvector matrix;

[0039] Using the whitening matrix R 1 and R 2 To perform whitening:

[0040] S 1 =PR 1 P T (4)

[0041] S 2 =PR 2 P T (5)

[0042] For S 1 and S 2 Perform eigenvalue decomposition:

[0043]

[0044] Matrix S 1 and S 2 The eigenvector matrix is ​​equal to, that is:

[0045] U 1 =U 2 =U 0 (8)

[0046] Then, the projection matrix W is:

[0047]

[0048] Select the first m and last m row vectors of the projection matrix to construct the spatial domain filter W m , the multi-channel EEG signal passes through the spatial domain filter to obtain a new signal:

[0049] z j =W m X (10)

[0050] Taking the logarithm and normalizing the variance, we get the eigenvector f:

[0051]

[0052] Where, j=1,…,2m; var represents the vector autoregression model.

[0053] Compared with the prior art, the present invention provides a motor imagery EEG decoding method based on adaptive optimal frequency band selection, which has the following beneficial effects:

[0054] Motor imagery-induced EEG signals have strong rhythmic characteristics, that is, there are significant changes in power spectrum energy in specific frequency bands. At the same time, this change has significant individual differences in the frequency domain. The present invention proposes a motor imagery EEG decoding method based on adaptive frequency band selection, which essentially selects the optimal signal frequency band from the perspective of the above-mentioned rhythmic changes. The traditional frequency band division method is mainly artificial, and its frequency band bandwidth and number of frequency bands are fixed. The present invention adopts a power spectrum peak frequency band selection method, and the frequency band bandwidth and number of frequency bands selected by this method have a high degree of individual adaptability. In summary, the present method has significant advantages and can effectively improve the accuracy of BCI classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is the structural block diagram mentioned in Embodiment 1 of the present invention;

[0056] Figure 2 The 22 lead distribution diagrams mentioned in Example 1 of the present invention;

[0057] Figure 3 This is a block diagram of the adaptive optimal frequency band selection method mentioned in Embodiment 1 of the present invention;

[0058] Figure 4 is the task-induced average power spectrum curve of the subject A01 mentioned in Example 1 of the present invention;

[0059] Figure 5 This is the task-induced average power spectrum curve of the subject A02 mentioned in Example 1 of the present invention. DETAILED DESCRIPTION

[0060] The present invention will be further described below in conjunction with specific embodiments, and the advantages and features of the present invention will become clearer as the description proceeds. However, these embodiments are exemplary only and do not constitute any limitation to the scope of the present invention. It should be understood by those skilled in the art that the details and forms of the technical solution of the present invention may be modified or replaced without departing from the spirit and scope of the present invention, but these modifications and replacements all fall within the scope of protection of the present invention.

[0061] It should be emphasized that, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the art to which the invention belongs. Although any methods, devices and materials similar or equivalent to those herein can be used in the practice or testing of the present invention, preferred methods, devices and materials are now described.

[0062] Embodiment 1:

[0063] See also Figure 1The present invention proposes a motor imagery EEG decoding method based on adaptive optimal frequency band selection, and its implementation process includes selecting key channels of motor imagery EEG signals and preprocessing them; calculating the power spectrum density of each channel signal, and using the power spectrum peak to adaptively select the optimal frequency band; using the selected frequency band to perform multi-band decomposition of the EEG signal, using CSP to extract spatial features from each sub-band data, and merging the features of each sub-band; using the mutual information optimal feature selection (MIBIF) algorithm to select the merged feature set; and using the SVM classifier for EEG decoding. The data source of the embodiment of the present invention is the public data set BCI Competition IV data set IIa. The data collected a total of 22 channel signals. The distribution diagram of the 22 channels is as follows Figure 2 As shown, the sampling frequency is 250 Hz; this embodiment uses the left / right hand motor imagery task data of subjects A01 and A02, and conducts a total of 12 groups of experiments, each group of left / right hand motor imagery tasks has 12 trials, therefore, a total of 288 trials (72 trials for each type of task); and the first 5 groups of experimental data are used as training sets, and the last group of experimental data is used as a test set.

[0064] The specific steps are as follows:

[0065] S1. Select key channels of motor imagery EEG signals and preprocess them;

[0066] The key channels of the EEG signals during the left and right hand motor imagery tasks are located in the motor sensory area. For this reason, the key channels selected in this embodiment are channels 2, 5, 7, 8, 9, 11, 12, 13, 14 and 18. Figure 2 shown.

[0067] The EEG signals of the selected key channels were preprocessed, including segmenting the EEG data, re-referencing using the common average reference (CAR), and removing power frequency interference and baseline drift.

[0068] S2, using adaptive optimal frequency band selection based on power spectrum peak for training set data: for the EEG data preprocessed in step S1, calculate the task-induced average power spectrum curve, and select the optimal frequency band based on the average power spectrum peak. Specifically, the following steps are included (such as Figure 3 shown):

[0069] (1) Construct a fifth-order Butterworth bandpass filter with a passband frequency of 8-36 Hz to perform bandpass filtering on the preprocessed EEG signal;

[0070] (2) Calculate the power spectral density of each channel signal using the Welch method;

[0071] (3) The power spectra of the key channels in the brain areas activated by different motor imagery tasks were averaged to obtain the task-induced average power spectrum curve;

[0072] (4) Smoothing the average power spectrum curve to remove pseudo peaks caused by small fluctuations in the curve;

[0073] (5) Obtaining the power spectrum peak value of the average power spectrum curve, where the number of peak values ​​is the number of frequency bands to be divided;

[0074] (6) Dynamically adjust and determine the bandwidth at each power spectrum peak.

[0075] Furthermore, the power spectra of the key channels of the brain regions activated by different motor imagery tasks in step (3) are averaged to obtain the task-induced average power spectrum curve; specifically: the key channels of the brain regions activated during the left-hand motor imagery task are located in the right brain region, namely channels 2, 7, 8, 9 and 14; the key channels of the brain regions activated during the right-hand motor imagery task are located in the left brain region, namely channels 6, 11, 12, 13 and 18. The power spectrum curves of channels 2, 7, 8, 9 and 14 during the left-hand motor imagery task and channels 6, 11, 12, 13 and 18 during the right-hand motor imagery task are averaged to obtain the task-induced average power spectrum curve:

[0076] like Figure 4 As shown, the A01 average power spectrum curve has two spectrum peaks, and their corresponding frequencies are 10.75Hz and 22.46Hz respectively.

[0077] like Figure 5 As shown, the A02 average power spectrum curve has three spectrum peaks, and their corresponding frequencies are 11.82Hz, 21.48Hz and 28.32Hz respectively.

[0078] Furthermore, step (6) dynamically adjusts and determines the bandwidth of each power spectrum peak, specifically comprising the following steps:

[0079] (a) For each power spectrum peak point obtained in step (5), the frequency value f corresponding to the peak point is 0 As the center, with H as the step length, determine the initial frequency band of the peak point [f 0 -H,f 0 +H]; the common spatial pattern algorithm (CSP) was used to extract the EEG features within the initial frequency band, and the support vector machine (SVM) classifier was used for feature classification to obtain the classification accuracy acc 0 ;

[0080] (b) With H as the step size, the bandwidth is increased to [f 0 -2*H,f 0+2*H], and use CSP to extract features, SVM to classify features, and obtain the classification accuracy acc 1 .

[0081] (c) If acc 1 ≥acc 0 , stop increasing the bandwidth, and the bandwidth used in step (b) is the bandwidth of the peak point; otherwise, repeat step (b) to increase the bandwidth, calculate the new classification accuracy, and update acc 1 .

[0082] (d) When the lower limit frequency f 0 -n*H≤8Hz or upper limit frequency f 0 When +n*H≥36Hz, the frequency band is stopped from increasing; at the same time, f 0 -n*H is 8Hz or f 0 +n*H is 36Hz, calculate the classification accuracy and set it as acc 1 ; if acc 1 ≥acc 0 , then the bandwidth at this time is the bandwidth of the peak point; otherwise, acc 0 The corresponding bandwidth is the peak point frequency band bandwidth.

[0083] Furthermore, in step (a), the frequency band step length H is determined by the number of peak points in the power spectrum. If the number of peak points is greater than 2, the step length H is 0.5 Hz; if the number of peak points is less than or equal to 2, the step length H is 1 Hz.

[0084] In this embodiment, the number of peak points of A01 is 2, and the step size H is 1 Hz. The optimal frequency bands determined by the adaptive optimal frequency band selection method based on the power spectrum peak are [8, 13.75] Hz and [20.46, 24.46] Hz. The number of peak points of A02 is 3, and the step size H is 0.5 Hz. The optimal frequency bands determined by the adaptive optimal frequency band selection method based on the power spectrum peak are [8, 15.82] Hz, [18.48, 24.48] Hz and [24.82, 31.82] Hz.

[0085] S3, EEG signal feature extraction and feature selection, specifically including the following steps: using the selected adaptive optimal frequency band, respectively decomposing the training set and test set EEG signals into multiple sub-bands; using the common spatial pattern (CSP) algorithm to extract features from each sub-band data; merging the multiple sub-band features of the training set data and the test set data into a training set feature vector and a test set feature vector, respectively; using the mutual information optimal feature selection (MIBIF) algorithm to select the best features to obtain the best features of the training set and the best features of the test set.

[0086] The common space pattern (CSP) feature extraction process in the application embodiment is as follows:

[0087] Assume X 1 and X 2 are two different types of data for motor imagery signals, and their covariance matrices are:

[0088]

[0089] In the formula, Respectively represent X 1 and X 2 , and trace(·) represents the trace of the matrix.

[0090] Perform eigendecomposition on the sum of the two covariance matrices R:

[0091] R=R 1 +R 2 =UΛU T (3)

[0092] Among them, Λ is the eigenvalue diagonal matrix of matrix R, and U is the eigenvector matrix.

[0093] Using the whitening matrix R 1 and R 2 To perform whitening:

[0094] S 1 =PR 1 P T (4)

[0095] S 2 =PR 2 P T (5)

[0096] For S 1 and S 2 Perform eigenvalue decomposition:

[0097]

[0098] Matrix S 1 and S 2 The eigenvector matrix is ​​equal to, that is:

[0099] U 1 =U 2 =U 0 (8)

[0100] Then, the projection matrix W is:

[0101]

[0102] Select the first m and last m row vectors of the projection matrix to construct the spatial domain filter W m , the multi-channel EEG signal passes through the spatial domain filter to obtain a new signal:

[0103] z j =W m X (10)

[0104] Taking the logarithm and normalization of the variance, we get the eigenvector f:

[0105]

[0106] Where, j=1,…,2m; var represents the vector autoregression model.

[0107] S4. Motor imagery feature classification: A support vector machine (SVM) classifier is used to classify features to achieve motor imagery pattern recognition. To verify the performance of this method, the EEG data collected from subjects A01 and A02 in the BCI Competition IV data set IIa were classified, and the classification accuracy of the test set samples was 89.35% and 98.37%, respectively, which reflects the good performance of this method.

[0108] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and improved concepts of the present invention within the technical scope disclosed by the present invention, and they should be covered by the protection scope of the present invention.

Claims

1. A motor imagery EEG decoding method based on adaptive optimal frequency band selection, characterized in that: The following steps are involved: S1. Key channel selection: Select key channels for motor imagery EEG signals and preprocess the selected key channels; S2, optimal frequency band selection: for the EEG data preprocessed in S1, the task-induced average power spectrum curve is calculated, and the optimal frequency band is adaptively selected according to the average power spectrum peak value; S3, EEG signal feature extraction and feature selection: Use the optimal frequency band selected in S2 to divide the training set and test set data into frequency bands; extract the features of each sub-band, and select the merged sub-band features; S4. Motor imagery feature classification: Use support vector machine classifier to classify features to achieve motor imagery pattern recognition.

2. The method for motor imagery EEG decoding based on adaptive optimal frequency band selection according to claim 1, characterized in that: The preprocessing described in S1 specifically includes segmenting the EEG data, re-referencing using a common average reference, and removing power frequency interference and baseline drift.

3. The method for motor imagery EEG decoding based on adaptive optimal frequency band selection according to claim 1, characterized in that: The S2 specifically includes the following contents: S2.1, construct a fifth-order Butterworth bandpass filter with a passband frequency of 8-36 Hz to perform bandpass filtering on the preprocessed EEG signal; S2.2, use the Welch method to calculate the power spectral density of each channel signal; S2.

3. Average the power spectra of key channels in brain regions activated by different motor imagery tasks to obtain the task-induced average power spectrum curve; S2.4, smoothing the average power spectrum curve to remove pseudo peaks caused by slight fluctuations in the curve; S2.5, obtaining the power spectrum peak value of the average power spectrum curve, wherein the number of peak values ​​is the number of frequency bands to be divided; S2.

6. Dynamically adjust and determine the bandwidth of each power spectrum peak.

4. The method for motor imagery EEG decoding based on adaptive optimal frequency band selection according to claim 3, characterized in that: The S2.3 further includes the following contents: the brain area activated by left-hand motor imagery is located in the right brain area, and the brain area activated by right-hand motor imagery is located in the left brain area; the power spectrum curves of all key channels of the right brain area during the left-hand motor imagery task and the left brain area during the right-hand motor imagery task are averaged to obtain the task-induced average power spectrum curve.

5. The method for motor imagery EEG decoding based on adaptive optimal frequency band selection according to claim 3, characterized in that: The S2.6 further includes the following: S2.6.

1. For each power spectrum peak point determined, with the frequency value f0 corresponding to the peak point as the center and H as the step length, determine the initial frequency band [f0-H, f0+H] of the peak point; use the common space pattern algorithm to extract the EEG features in the initial frequency band, use the support vector machine classifier to perform feature classification, and obtain the classification accuracy acc0; S2.6.

2. With a step size of H, increase the frequency band width to [f0-2*H, f0+2*H], use the common space pattern algorithm to extract features, and use the support vector machine classifier to classify the features to obtain the classification accuracy acc1. S2.6.3, if acc1 ≥ acc0, stop increasing the frequency band, and the bandwidth used in S2.6.2 is the frequency band bandwidth of the peak point; Otherwise, repeat S2.6.2 to increase the bandwidth, calculate the new classification accuracy, and update acc1; S2.6.

4. When the lower limit frequency of the frequency band f0-n*H≤8Hz or the upper limit frequency f0+n*H≥36Hz, stop increasing the frequency band; at the same time, determine f0-n*H to be 8Hz or f0+n*H to be 36Hz, calculate the classification accuracy, and set it as acc1; if acc1≥acc0, the bandwidth at this time is the bandwidth of the peak point; otherwise, the bandwidth corresponding to acc0 is the bandwidth of the peak point frequency band.

6. The method for motor imagery EEG decoding based on adaptive optimal frequency band selection according to claim 5, characterized in that: The frequency band step length H is determined by the number of peak points of the power spectrum. If the number of peak points is greater than 2, the step length H is 0.5 Hz; if the number of peak points is less than or equal to 2, the step length H is 1 Hz.

7. The method for motor imagery EEG decoding based on adaptive optimal frequency band selection according to claim 1, characterized in that: The S3 specifically includes the following contents: S3.1, using the selected adaptive optimal frequency band, decompose the EEG signals of the training set and the test set into multiple sub-bands respectively; S3.2, using the common spatial pattern algorithm to extract features from each sub-band data; S3.3, merging multiple sub-band features of the training set data and the test set data into a training set feature vector and a test set feature vector respectively; S3.

4. Use the mutual information optimal feature selection algorithm to select the best features and obtain the best features of the training set and the best features of the test set.

8. The method for motor imagery EEG decoding based on adaptive optimal frequency band selection according to claim 7, characterized in that: Extracting the features of each sub-band in S3.2 specifically refers to extracting features from each sub-band data using a common spatial pattern algorithm, which specifically includes the following contents: Assume that X1 and X2 are two different types of data of motor imagery signals, and their covariance matrices are: In the formula, denote the transpose of X1 and X2 respectively, and trace(·) denotes the trace of the matrix; Perform eigendecomposition on the sum of the two covariance matrices R: R =R1+R2=UΛU T (3) Among them, Λ is the eigenvalue diagonal matrix of matrix R, and U is the eigenvector matrix; Using the whitening matrix Whiten R1 and R2: S1=PR1P T (4) <h2 style=";text-align:left;direction:ltr">S2=PR2P<h2 style=";text-align:left;direction:ltr"> T <h2 style=";text-align:left;direction:ltr"> (5) Perform eigenvalue decomposition on S1 and S2: The eigenvector matrices of matrices S1 and S2 are equal, that is: U1=U2=U0 (8) Then, the projection matrix W is: Select the first m and last m row vectors of the projection matrix to construct the spatial domain filter W m , the multi-channel EEG signal passes through the spatial domain filter to obtain a new signal: z j =W m X (10) Taking the logarithm and normalizing the variance, we get the eigenvector f: Where, j=1,…,2m; var represents the vector autoregression model.