CN112528834B - Electroencephalogram signal cross-test classification method of subband target alignment co-space mode - Google Patents

Abstract

The invention discloses a sub-band target alignment co-space mode electroencephalogram signal cross-test classification method. The invention firstly carries out band-pass filtering on the electroencephalogram signals into a plurality of sub-band signals (sub-band filtering), then adopts CSP algorithm to extract characteristics, then selects more representative characteristics through a minimum redundancy maximum correlation method, and finally uses a traditional linear discriminant analysis classifier for electroencephalogram classification. The invention combines subband filtering with target alignment to effectively improve the performance of cross-test classification.

Description

Electroencephalogram signal cross-test classification method of subband target alignment co-space mode

Technical Field

The invention belongs to the field of pattern recognition, and relates to a feature extraction and cross-test classification method for acquiring more frequency band information and reducing domain distribution difference by a target alignment method by utilizing subband filtering aiming at an electroencephalogram signal.

Background

Brain-computer interface (BCI) is a man-machine interaction system based on brain signals, and is widely applied to various fields such as exoskeleton rehabilitation robots, fatigue detection, intelligent home, entertainment games and the like. In most BCIs, electroencephalogram (EEG) is the most commonly used input signal due to its non-invasive and inexpensive advantages compared to signals such as brain magnetic imaging (MEG), functional magnetic resonance imaging (fMRI), and Positron Emission Tomography (PET). Among them, motor Imagery (MI) is a popular experimental example. However, EEG is a very weak electrical signal acquired from the scalp, which has poor spatial resolution, low signal-to-noise ratio, and significant differences between subjects and stages, which makes analysis and identification of EEG signals very difficult. Thus, it becomes very interesting to study efficient methods to extract and identify the brain electrical features of different tasks.

In an electroencephalogram-based brain-computer interface system, decoding different electroencephalograms needs to be subjected to complex signal processing, feature extraction and classification. Co-space model (CSP) is a very popular and used feature extraction method. Its performance depends on the specific frequency band of the EEG, and classification using CSP features is less accurate when the EEG signal is filtered using an inappropriate frequency range. To address this problem, novi and Guan propose a subband co-spatial mode (SBCSP) that uses a set of filters to divide the EEG into multiple subbands, then inputs the extracted features into the LDA to obtain scores representing the classification capabilities of each subband, and finally fuses the scores to make a decision. Ang and Guan propose a filter bank co-space pattern (FBCSP) algorithm and devised various feature selection methods to select more representative features. The two methods carry out subband filtering on the original electroencephalogram data, new characteristic representation is obtained through fusion and characteristic selection, and finally a better experimental result is obtained. Obviously, subband filtering becomes an effective electroencephalogram signal processing method.

However, the SBCSP and FBCSP methods perform much more satisfactorily in cross-stage classification than in cross-test classification because individual differences between users are much greater than differences between different stages of the same user. In recent years, more and more researchers introduce ideas of transfer learning into electroencephalogram classification to solve the problem of data distribution adaptation. Transfer learning utilizes data/information from one or more source domains to improve learning performance of a target domain species and has been successfully applied in BCI. Zanini et al propose a Riemann Alignment (RA) method to align the EEG covariance matrices of different subjects, and the aligned covariance matrices can be classified as features directly using an MDRM classifier. Yair et al propose a method for domain adaptation using parallel transmission on the cone-flow rows of a symmetric positive definite matrix. He and Wu extend the RA method to the Euclidean Alignment (EA) method to align EEG trials from different subjects in euclidean space making them more uniform so that any euclidean spatial classifier can be used after it. Zhang and Wu propose a new popular embedded knowledge Migration (MEKT) method that first extracts tangential spatial features by RA versus Ji Xie variance matrices, then performs domain adaptation by minimizing joint probability distribution shifts between source and target domains, while preserving their geometry.

Disclosure of Invention

Because the performance of CSP depends on the working frequency band and the difficulty of cross-test classification is great, the invention provides a characteristic extraction method based on sub-band filtering and field alignment, which is called sub-band target alignment co-space mode (SBTACSP), and designs an electroencephalogram cross-test classification method of sub-band target alignment co-space mode based on the mode.

The invention comprises the following steps:

step (1): and (5) pretreatment. Collecting multichannel motor imagery brain point signals of a subject; then, the electroencephalogram signal in a specific frequency range is subjected to subband filtering, and the electroencephalogram signal is subjected to bandpass filtering in a predetermined bandwidth to obtain a plurality of subband signals.

Step (2): a new field alignment method is provided, namely, target alignment, which aligns a source field (test) sample into a target field (training) space and applies the source field (test) sample to signals of each frequency band, only the source field sample is aligned and changed, the target field sample is not changed, and then the target field sample is extracted for electroencephalogram characteristics.

Step (3): and filtering the aligned electroencephalogram signals of each frequency band by using CSP, and then splicing the CSP characteristics of a plurality of frequency bands together to form a CSP characteristic vector with higher dimension.

Step (4): the p most representative features are selected to form the final feature according to a minimum redundancy maximum correlation (mRMR) feature selection algorithm.

Step (5): and classifying the features with the number p of the features after feature selection by adopting an LDA classifier.

Compared with the FBCSP and EA-CSP-LDA methods, the invention has the following advantages:

compared with a filter bank co-space mode (FBCSP) method, the method provided by the invention has the advantages of utilizing the advantages of transfer learning in cross-test classification, reducing inter-domain distribution difference, obtaining more representative characteristics for classification and improving classification accuracy. Compared with the Euclidean alignment method, the method adopts the sub-band filtering processing method to obtain more frequency band information, thereby improving the classification performance. Compared with the sub-band Euclidean alignment co-space mode, the new field alignment method provided by the invention, namely target alignment, is more suitable for being combined with sub-band filtering, and the obtained classification precision is better.

Description of the drawings:

FIG. 1 is a flow chart of the method of the present invention;

FIG. 2 (a) shows the result of 10-fold cross-validation of parameter p for BCI composition IV Dataset IIa;

FIG. 2 (b) shows the result of 10-fold cross-validation of parameter p for BCI composition IV Dataset IIb;

FIG. 3 (a) is a characteristic t-SNE visualization effect graph of two subjects CSP and SBTACSP in Dataset Dataset IIa in a multi-source domain classification;

FIG. 3 (b) is a characteristic t-SNE visualization effect graph of two subjects CSP and SBTACSP in Dataset Dataset IIb in a multisource domain classification.

Detailed Description

The present invention is described in detail below with reference to the accompanying drawings.

In the field of brain-computer interfaces, the classical co-spatial mode (Common spatial pattern, CSP) approach requires efficient selection of specific frequency bands, while the high instability and individual variability of the electroencephalogram signals increases the difficulty across the classification under test. In order to solve the problems, the invention provides a subband target alignment co-space mode (Sub-Band Target Alignment CSP, SBTACSP) method, and applies the method to cross-test classification of brain electrical signals, and fig. 1 is a flow chart of the invention.

As shown in fig. 1, the implementation of the method of the present invention mainly comprises 5 steps: (1) collecting multichannel brain electrical signals and sub-band filtering; (2) performing target alignment separately for each subband; (3) Carrying out co-space mode feature extraction on each sub-band after alignment, and splicing the sub-bands into a feature vector with higher dimension; (4) Selecting the most representative p feature composition final features through mRMR; (5) And inputting the extracted most representative features into a classifier for classification to obtain a result.

The steps are described in detail one by one.

Step (1): and (5) data acquisition. Selecting BCI Competition public data BCI composition IV data IIa: the dataset consisted of electroencephalogram data of 9 healthy subjects. The data for each subject consisted of an electroencephalogram of four motor imagery of the left hand, right hand, foot and tongue of one subject. Each experimental signal was recorded using 25 electrodes, the electrode positions using the international 10/20 system. The 22 channel EEG signal and the 3 channel EOG signal were recorded and sampled at 250 Hz. In this example, only two classes (left and right) of data were selected for testing, 72 trials for each class. BCI composition iv Dataset iib: the dataset consisted of electroencephalogram data of 9 healthy subjects as well. Its test was performed with only left and right hand movements. Each experimental signal was recorded using 6 electrodes, the electrode positions using the International 10/20 System. The EEG signals of 3 channels and the EOG signals of 3 channels were recorded and sampled at 250 Hz.

Sub-band filtering: subjects of all data sets were subjected to the same subband filtering process. Each subject selects EEG motor imagery data of the subject in a time period of 0.5-3.5s after visual prompt, and then filters the data in a frequency range of 8-32Hz by using a 50-order FIR filter with 4Hz as a bandwidth to obtain 6 sub-bands, namely 8-12Hz,12-16Hz, … and 28-32Hz.

Step (2): target alignment is performed separately for each subband. The method specifically comprises the following steps: giving an N-channel space-time EEG signal matrixWhere T represents the number of samples per channel. First, calculating target alignment to obtain a reference matrix:

wherein n is _t Indicating the total number of target domain t trials,the ith trial signal for the kth subband of the target domain,>the Euclidean average of all trials for the kth subband of the target domain.

The Euclidean average of all experiments of the kth subband of the source domain is then obtained using the same calculation method as in equation (1)Finally, n of the kth sub-band of the source domain is processed by the method (2) _s The trial was performed for target alignment.

Wherein,the ith test signal for the kth subband of the source domain,>the signal after target alignment is tested for the kth subband i time in the source domain.

Step (3): and carrying out CSP feature extraction on each sub-band after alignment, and splicing the CSP feature extraction into a high-dimensional feature vector. The method specifically comprises the following steps: firstly, linear transformation is carried out on the aligned electroencephalogram signals by using a CSP algorithm, and spatial filtering is realized:

wherein,is the CSP projection matrix of the kth subband. Spatially filtered signal Z in (3) _k,i Use->Thereby maximizing variance differences between the two classes of bandpass filtered EEG. These two classes may include left and right hand motor imagery data, with the m pairs of CSP features for the kth subband, the ith trial, defined as:

wherein the method comprises the steps of Represents +.about.m for the first and last m columns>diag (·) represents the diagonal elements of the matrix, tr (·) represents the sum of the diagonal elements of the matrix. Thus, the SBTACSP characteristic of the ith trial can be expressed as:

v _i ＝[v _1,i ,v _2,i ,...,v _6,i ] (5)

wherein the method comprises the steps ofThus, the training data from the source domain can be characterized as +.>

Step (4): the p features that are most representative are selected by mRMR to make up the final feature. According to the feature extraction in the step (3), the features of the source domain (training) data can be obtainedAnd real tag->

This makes the feature dimension too large, since the feature of one trial is stitched from 6 sub-band features. In order to reduce the feature dimension, the difficulty of learning tasks is reduced, and the efficiency of the model is improved. From each f using an mRMR feature selection algorithm _i And selecting the most representative CSP features from the features. This can be achieved by maximizing the following expression:

wherein n is _f =6×2m, i (·) is mutual information, r _i The value of the i-th feature mRMR is expressed. For r _i Sorting, screening out the features corresponding to p maximum r values, and finally obtaining a source domain (training) feature matrix

Step (5): the obtained source domain feature matrixThe method is used for inputting a training model into an LDA classifier, testing a target domain to obtain a test result, and verifying the classification performance of the method in electroencephalogram signal feature extraction and cross-test classification.

To verify the effectiveness of the method of the present invention, an assessment will be made in two scenarios across the test, including single source to Single Target (STS) migration and multiple source to single target (MTS) migration. Meanwhile, compared with the performance of methods such as MDRM, RA-MDRM, CSP-LDA, EA-CPS-LDA, FBCSP, SBEACSP and the like on the BCI composition IV data set IIa and IIb, the feasibility of the invention is measured by testing the classification accuracy.

Increasing m does not significantly improve classification accuracy according to the CSP method. For data set IIa, parameter m for CSP feature extraction was set to 3, so the number of preliminary features extracted per trial was 36 (6X 6). For data set IIb, since the data set uses only 3 EEG channels, and in equation (3)The maximum choice of m is limited, so the parameter m is set to 1, and the number of preliminary features extracted per trial is 12 (2×6).

Fig. 2 (a) and 2 (b) show the experimental results of 10 fold cross-validation of two datasets after SBTACSP selects different feature numbers p. As can be seen from fig. 2 (a), the accuracy of the cross-validation is highest when the number of features is 18, and therefore, the p-value is 18 optimal for the data set iia. As can be seen from fig. 2 (b), the result of the cross-validation is optimal when the number of features is 12, and therefore, the p-value is 12 optimal for the data set iib.

Table 1 gives the results of the accuracy of the 10-fold cross-validation of the proposed method with 3 methods on both data sets, based on the optimal p-values obtained in fig. 2 (a) and fig. 2 (b). As can be seen from Table 1, the average results of the SBTACSP method are better than RA-MDRM, EA-CSP-LDA and FBCSP in both dataset IIa and IIb, and are optimal over more than half of the subjects, with better results in dataset IIa for subjects 3, 5, 6, 7, 8 and 9. In dataset IIb, subjects 3, 4, 5, 6 and 8 performed better. While SBTACSP is preferred over FBCSP because the data is aligned before CSP extracts features, SBTACSP is preferred over EA-CAP-LDA because the 8-32Hz band is sub-band processed, and the latter three methods are preferred over RA-MDRM because real labels are used in the test. These also illustrate that the SBTACSP method is feasible and that the choice of parameter p is reasonable and efficient.

TABLE 1 accuracy results of 10 fold Cross-validation of RA-MDRM, EA-CSP-LDA, FBCSP and SBTACSP on BCI completions IV data IIa and IIb

Tables 2 and 3 show the classification accuracy of BCI completions iv data iia and iib on STS and MTS migration, respectively. In general, SBTACSP is superior to the other five methods in terms of classification accuracy. Specifically, for dataset 2a, in STS migration, the present invention achieved optimal average accuracy over subjects a01 and a06, SBEACSP achieved optimal results over a03, FBCSP performed best over a02 and a07, RA-MDRM performed best over a04, a05, a08 and a 09; in MTS migration, the present invention achieved optimal accuracy in subjects A01, A04, A06 and A07, while FBCSP performed optimally in A02 and A09 and EA-CSP-LDA performed best in A03 and A08. For dataset 2B, the present invention performed optimally on subjects B01, B03, B04, B05, B08, and B09, with SBEACSP achieving the best average accuracy outside of B02, B06, and B07 during STS migration. In MTS migration, SBTACSP achieved the best accuracy on subjects B02, B04, B06 and B09, SBEACSP performed best on B05, and FBCSP performed best on B03, B08. Taking MTS migration as an example, the classification accuracy of the present invention in subject A01 was 88.19%, better than EA-CSP-LDA (86.81%) and RA-MDRM (73.61%), the average accuracy of the method was highest (75%) among all nine subjects, 7.29% higher than FBCSP, 2.55% higher than RA-MDRM, 1.39% higher than SBEACSP, and 1.23% higher than EA-CSP-LDA. Compared to STS and MTS migration, MTS migration has much higher accuracy than STS migration, which suggests that multi-source domain migration may improve accuracy properly.

In the cross-test scene, the FBCSP does not achieve a good effect, and the effect in the cross-stage classification is far better than that of the cross-test classification shown, because the data of the training set and the test set of the cross-stage classification belong to the same person, but the acquired time is different, and the difference degree of the training and the test data is far less than that of the cross-test. Except for the STS migration condition of the data set 2b, the average classification accuracy of the FBCSP is higher than that of the CSP-LDA, which shows that the band division filtering achieves positive effects. At the same time, the classification accuracy of the EA-CSP-LDA and SBTACSP, SBTACSP is at least 1.2% higher than that of the EA-CSP-LDA, and the maximum difference can reach 5.22%, which also indicates that subband filtering before alignment is necessary. Comparing FBCSP with SBTACSP, SBTACSP performed better, indicating that TA could effectively solve the problem of data variability between subjects. Meanwhile, the EA-CSP-LDA classification effect is also superior to that of the FBCSP. These descriptions have great prospects for development across the great advantages of transfer learning in the class under test. Combining the above three points, it is possible to demonstrate that subband filtering and target alignment are combined. Finally, better than SBEACSP, SBTACSP, which illustrates that the proposed variant of the invention is more suitable in combination with subband filtering.

TABLE 2 Classification accuracy (%)

TABLE 3 Classification accuracy (%)

Fig. 3 (a) and (b) visualize the classification results using a nonlinear dimension reduction technique t-SNE to compare the differences between features extracted using CSP and those extracted using SBTACSP. FIG. 3 (a) shows the t-SNE visualization of subjects A01 and A03 in dataset IIa in MTS migration, each row corresponding to a different test object, with Xs1 and Xs2 representing two classes of features of the source domain, and Xt1 and Xt2 representing two classes of features of the target domain, respectively. The left plot shows features of direct CSP extraction of EEG data and the right plot shows features of SBTACSP. The results of the visualization of subjects B04 and B09 in dataset iib are shown in fig. 3 (B). By comparison, it can be found that the source (training) and target (testing) domain features of the SBTACSP overlap each other, i.e. the differences between them are reduced, while the distributions between the same class in different domains overlap as much as possible. It can thus be seen that the visualization effect is consistent with the classification result. This also demonstrates the feasibility of the invention.

In summary, the invention provides a subband target alignment co-space mode motor imagery electroencephalogram signal cross-test classification method, more frequency band information is obtained through a subband filtering signal processing method, and meanwhile, an unsupervised target alignment method is utilized to reduce the distribution difference of a source domain and a target domain. The invention combines the advantages of subband filtering and field alignment, and improves the feature and the accuracy of cross-test classification.

Claims (4)

1. The subband target alignment co-space mode electroencephalogram signal cross-test classification method is characterized by comprising the following steps of: the method comprises the following main steps:

step (1): the data acquisition selects BCI Competition public data, and comprises two data sets, namely a BCI composition IV data set IIa and a BCI composition IV data set IIb;

sub-band filtering: subjecting subjects of all data sets to the same sub-band filtering process; each subject selects EEG motor imagery data of the subject in a time period of 0.5-3.5s after visual prompt, and filters the data in a frequency range of 8-32Hz by using a 50-order FIR filter with 4Hz as a bandwidth to obtain 6 sub-bands, namely 8-12Hz,12-16Hz, … and 28-32Hz;

step (2): aligning the source domain sample into the target domain space by adopting a domain alignment method, namely aligning the target, namely aligning the test sample into the training space;

the method specifically comprises the following steps: giving an N-channel space-time EEG signal matrixWhere T represents the number of samples per channel, and calculating the target alignment to obtain a reference matrix:

wherein n is _t Indicating the total number of target domain t trials,the ith trial signal for the kth subband of the target domain,>euclidean means for all trials for the kth subband of the target domain;

next, the Euclidean average of all experiments of the kth subband of the source domain is obtained by using the same calculation method as the formula (1)

Finally, n of the kth sub-band of the source domain is processed by the method (2) _s Performing target alignment for the secondary test;

wherein,the ith test signal for the kth subband of the source domain,>the signal after the ith test target alignment of the kth subband in the source domain;

step (3): filtering the aligned electroencephalogram signals of each frequency band by using a CSP algorithm, and then splicing CSP features of a plurality of frequency bands together to form a feature vector with higher dimension;

step (4): selecting p most representative features to form final features according to a minimum redundancy maximum correlation feature selection algorithm;

step (5): the obtained source domain feature matrixInputting a training model in the LDA classifier, and testing a target domain to obtain a test result;

wherein the step (4) specifically comprises: according to the feature extraction in the step (3), the features of the source domain data are obtainedAnd real tag->

From each f using an mRMR feature selection algorithm _i The best represented CSP feature is selected from the features, and is realized by maximizing the following expression:

wherein n is _f =6×2m, i (·) is mutual information, r _i A value representing the measure of the ith feature mRMR;

for r _i Sorting, screening out the features corresponding to p maximum r values, and finally obtaining a source domain feature matrix

2. The subband target alignment co-space mode electroencephalogram signal cross-test classification method according to claim 1, wherein the method comprises the following steps of: the BCI composition IV Dataset IIa in the step (1) consists of the brain electrical data of 9 healthy subjects; the data of each subject consists of electroencephalograms of four motor imagination of the left hand, the right hand, the foot and the tongue of one subject; each experimental signal is recorded by using 25 electrodes, and the positions of the electrodes adopt an international 10/20 system; the 22 channel EEG signal and the 3 channel EOG signal were recorded and sampled at 250Hz, while only the left and right hand data were selected for testing, 72 trials each.

3. The subband target alignment co-space mode electroencephalogram signal cross-test classification method according to claim 1, wherein the method comprises the following steps of: BCI composition iv Dataset ii b consisted of electroencephalogram data of 9 healthy subjects; its test only performs left-right hand movements; each experimental signal is recorded by using 6 electrodes, and the positions of the electrodes adopt an international 10/20 system; the EEG signals of 3 channels and the EOG signals of 3 channels were recorded and sampled at 250 Hz.

4. The subband target alignment co-space mode electroencephalogram signal cross-test classification method according to claim 1, wherein the method comprises the following steps of: the step (3) is specifically as follows: the CSP algorithm is utilized to carry out linear transformation on the aligned electroencephalogram signals, and spatial filtering is realized:

wherein,is the CSP projection matrix of the kth subband;

the m pairs of CSP features for the kth subband, ith trial, are defined as:

wherein the method comprises the steps ofRepresents +.about.m for the first and last m columns>diag (·) represents the diagonal elements of the matrix, tr (·) represents the sum of the diagonal elements of the matrix;

the SBTACSP characteristic of the ith trial was expressed as:

v _i ＝[v _1,i ,v _2,i ,...,v _6,i ] (5)

wherein the method comprises the steps of

The training data from the source domain is characterized as