CN109977810A

CN109977810A - Brain electricity classification method based on HELM and combination PTSNE and LDA Fusion Features

Info

Publication number: CN109977810A
Application number: CN201910183172.XA
Authority: CN
Inventors: 段立娟; 连召洋; 乔元华; 陈军成
Original assignee: Beijing University of Technology
Current assignee: Beijing University of Technology
Priority date: 2019-03-12
Filing date: 2019-03-12
Publication date: 2019-07-05
Anticipated expiration: 2039-03-12
Also published as: CN109977810B

Abstract

The invention discloses a motor imagery electroencephalogram classification method based on HELM and combined with PTSNE manifold and LDA feature fusion, and improves the classification accuracy. In terms of feature extraction, on the one hand, using PCA combined with LDA method to extract linear features can not only eliminate noise, but also consider the label information of training data; Complex nonlinear intrinsic manifold features. In terms of feature classification, HELM algorithm with high classification accuracy is used for classification and recognition of motor imagery EEG signals.

Description

Electroencephalogram classification method based on HELM and combined PTSNE and LDA feature fusion

Technical Field

The invention belongs to the technical field of electroencephalogram signal processing, and particularly relates to a motor imagery electroencephalogram classification method based on HELM and combining PTSNE manifold and LDA feature fusion.

Background

The brain electrical signal of the motor imagery is used for identifying the brain state, and the brain imagery of the human can be controlled according to the brain electrical signal. The analysis of the brain electrical signals is greatly helpful for patients with encephalopathy. Because the acquired electroencephalogram signal is a non-stationary signal which has strong randomness and visually lacks regularity in waveform, an effective feature extraction method is required to improve the classification accuracy of the electroencephalogram signal.

The feature extraction method commonly used in electroencephalogram analysis includes time domain, frequency domain and time-frequency domain combination. In the time domain, the peak detection is highly targeted, but also highly subjective. The autoregressive model, while not a priori known and relatively less subjective, is relatively sensitive to artifact signals. In the frequency domain, the power spectral density may reflect the energy variation, but substantially loses the time domain information. In time-frequency domain combining, the wavelet transform can obtain the waveform of the target frequency band, but it requires a priori knowledge of the target frequency band.

In the aspect of feature extraction, in consideration of higher complexity, lack of regularity of waveforms of electroencephalogram signals and sparsity of signals, original data needs to be clustered and subjected to dimension reduction. Although PCA can reduce the dimension of the original data according to the accumulated contribution rate, PCA is an unsupervised method and ignores the label information of the data. And LDA is supervised learning, and fully considers the label information of the training data. The data points may be clustered according to categories in the low-dimensional space, with data points of the same category being closer together in the low-dimensional mapping space. However, both methods are linear transformations, and neither takes into account the non-linear manifold structure inherent in the brain electrical signal. Common non-linear manifold methods such as ISOMAP, Laplace eigenmap, local neighborhood embedding, random neighborhood embedding, and T-distributed random neighborhood embedding (TSNE). TSNE tries to maintain the probability distribution of the original data in the process of mapping the high-dimensional original data to the low-dimensional embedded coordinates. PTSNE learns the mapping between gaussian distributions in high dimensional space and student T distributions in low dimensional space by minimizing KL divergence through a neural network. PCA, LDA and PTSNE have advantages and disadvantages, and each algorithm is single and practical and cannot well solve the problem of electroencephalogram feature extraction.

Disclosure of Invention

Aiming at the background, the invention provides a motor imagery electroencephalogram classification method based on HELM and combining PTSNE manifold and LDA feature fusion, and the classification accuracy is improved. In the aspect of feature extraction, on one hand, linear features are extracted by combining PCA with LDA, so that noise can be eliminated, and label information of training data can be considered; on the other hand, the nonlinear combination characteristics are obtained through PTSNE and LDA, and the complex nonlinear internal manifold characteristics in the brain electricity can be discovered. In the aspect of feature classification, an HELM algorithm with high classification accuracy is adopted for classification and identification of the motor imagery electroencephalogram signals.

In order to achieve the purpose, the invention adopts the following technical scheme:

a motor imagery electroencephalogram classification method based on HELM and combined with PTSNE manifold and LDA feature fusion comprises the following steps:

step (1) preprocessing of EEG signals

The obtained electroencephalogram data are randomly disturbed and normalized, and the data are segmented by overlapped sliding time windows to obtain segmented subdata.

Step (2) feature extraction and fusion

a. Linear feature extraction

The first pass obtains linear binding characteristics through PCA and LDA.

1) The first path firstly carries out PCA dimension reduction on each segment of segmented data respectively.

2) Secondly, LDA conversion is carried out on each segment of segmented data after PCA dimension reduction.

3) Linear binding characteristics after PCA and LDA were obtained.

b. Nonlinear feature extraction

The other way obtains linear binding characteristics through PCA and LDA.

1) And the other path firstly carries out PTSNE dimension reduction on each segment of segmented data respectively.

2) Secondly, LDA conversion is carried out on each segment of segmented data after PTSNE dimension reduction.

3) Nonlinear binding characteristics after PCA and LDA were obtained.

c. Two-way feature fusion

And performing feature fusion on the obtained linear combination features and nonlinear combination features to serve as finally extracted features.

Step (3) feature classification

In terms of classification method, the HELM is a deep classification model, which can be divided into two parts: ELM SAE is used as feature extraction and the underlying ELM is used for feature classification. The classification accuracy is improved, and the first-order normal form constraint is added, so that the algorithm is simpler and more sparse.

Drawings

FIG. 1 is a block diagram of a process according to the present invention;

Detailed Description

The invention is further described with reference to the accompanying drawings and the detailed description.

The process of the method comprises the following steps:

(1) and (4) preprocessing the electroencephalogram signals.

Firstly, the obtained electroencephalogram data are randomly disordered and normalized. Then, considering the complexity and instability of the brain electrical signal, we use overlapping sliding window segments to preserve the useful information in the brain electrical signal. Based on early work in the laboratory, dominant electrodes a1 and a2 were also selected, each with 896 dimensions. The data for each electrode is divided into 9 segments by a 500ms time window and a 125ms overlapping window, each data segment having 128 dimensions.

(2) Feature extraction and fusion

The data segments divided in the a1 and a2 electrodes were copied into two equal parts, each of which was characterized by a different method.

a. Linear feature extraction

The first pass obtains linear binding characteristics through PCA and LDA.

Obtaining the weight W by optimizing an objective function L (W)

L(w)＝w^TCw-λ(w^Tw-1)

Where C is the covariance of the input sub-data segment X.

Obtaining Y from the subdata X and the obtained W

Y＝XW

Wherein,

and performing PCA on each subdata in 18 data segments of the first part of data and reducing the D to 16 dimensions, wherein the accumulative contribution rate is greater than 99%.

Obtaining the weight w by optimizing an objective function L (w)

L(w)＝w^TS_Bw-λ(w^TS_Ww-1)

Wherein S_WAnd S_BRespectively by segmentation after PCAThe intra-class and inter-class covariances of the sub-feature X.

Obtaining output characteristic Y of LDA by inputting sub-characteristic X and obtained W

Y＝W^TX

An LDA operation is performed on each subdata of 18 data segments of 561 samples, respectively. Since the experimental dataset is a binary problem, each 16-dimensional feature subdata is mapped to a 1-dimensional space.

3) Linear binding characteristics after PCA and LDA were obtained.

Each sample has 18 1-dimensional feature sub-data, and the 18 1-dimensional feature sub-data are combined into an 18-dimensional feature vector, and the 18-dimensional feature vector is a linear combination feature.

b. Nonlinear feature extraction

The mapping between the gaussian distribution in the high dimensional space and the student T distribution in the low dimensional space is learned by minimizing the KL divergence through a neural network.

The KL divergence objective function is as follows:

wherein q is_ijIs the probability distribution of data in a low dimensional space, which is defined as follows:

p_ijis the probability distribution of data in a high dimensional space, defined as follows:

wherein p is_j|iIs the probability of sample j under sample i. X and Y are the high-dimensional input data and the low-dimensional output features of the segmented sub data, respectively. Sigma²Is the variance.

The PTSNE operation is first performed on each subdata in the 18 data segments of the second share of data. Because of the limited boltzmann machine inside the PTSNE, the PTSNE needs to be trained first, and then features in the test data are extracted using the trained PTSNE. When the dimension is 30, the classification performance is better. Thus, after PTSNE extraction features, there are 18 data segments per sample, and each data segment has 30-dimensional features.

Subsequently, an LDA operation is performed on each subdata of 18 data segments of the 561 samples, respectively, and each 16-dimensional feature subdata is mapped to a 1-dimensional space.

3) Nonlinear binding characteristics after PCA and LDA were obtained.

Each sample has 18 1-dimensional feature sub-data, the 18 1-dimensional feature sub-data are combined into an 18-dimensional feature vector, and the 18-dimensional feature vector is a nonlinear combined feature.

c. Two-way feature fusion

And finally, generating linear combination features by the first path and generating nonlinear combination features by the second path, fusing the two paths of combination features together, and taking the fused combination features as final features of feature extraction and input features of the HELM.

(3) Feature classification

In terms of classification method, the HELM is a deep classification model, which can be divided into two parts: ELM SAE is used as feature extraction and the underlying ELM is used for feature classification. The method improves the classification accuracy and increases the first-order normal form constraint to make the algorithm more concise and sparse. Therefore, the HELM algorithm with high classification accuracy is selected as a classification model of the electroencephalogram signals.

The objective function of HELM is as follows:

β＝argmin{||Hβ-X||²+||β||_l1}

where β is the hidden layer weight, l1 is a first order normal constraint.

Wherein A is_iAnd b_iRespectively a weight matrix and an offset initialized at random. k is the number of layers of the ELM combined with sparse self-encoding.

The hidden layer weight of the last layer of the HELM, i.e. the base ELM, is:

where C is the regularization coefficient and I is the identity matrix T which is the true label value.

For the HELM class model, the regularization coefficient C ranges from 10^-10，10^-9，…，10⁹，10¹⁰. The first and second tier hidden tier nodes L1 are arranged in a range of 10, 20, …, 500. The third hidden layer node L2 is set to range from 100, 200, …, 1000. To verify the stability of the algorithm, the algorithm was run 50 times under the same parameters, and the average classification accuracy and the highest classification accuracy among these 50 were calculated.

Comparative analysis

Table 1 shows the comparison of the accuracy of the method of the present invention with other electroencephalogram classification algorithms, and Table 2 shows the comparison of the accuracy of the method of the present invention with the mainstream electroencephalogram classification algorithm.

TABLE 1 comparison of the accuracy of the method of the invention with other algorithms

TABLE 2 comparison of the accuracy of the method of the present invention with the mainstream electroencephalogram classification algorithm

A second BCI competition public data set is adopted in the comparison experiment to verify the classification performance of the feature fusion algorithm. The experimental combinations are shown in tables 1 and 2, where the values in the parenthesis represent the highest classification accuracy for 50 runs of the algorithm under the same parameters, the values before + -represent the mean classification accuracy for 50 runs, and the values after + -represent the variance of the classification accuracy for 50 runs. The second column is the name of the feature extraction method employed, and the third column is the name of the classification model employed. Firstly, PTSNE is only used for feature extraction, SVM is used as a classification model, and the classification accuracy can reach 90.785%. Then, PCA or LDA is used for feature extraction, HELM is used as a classification model, and the classification accuracy rate respectively reaches 91.81% and 84.64%. PTSNE is combined with LDA to be used for feature extraction, HELM is used as a classification model, and the classification accuracy can reach 93.174%. In addition, PCA is combined with LDA to be used for feature extraction, HELM is used as a classification model, the highest classification accuracy rate can reach 94.54%, and the average classification accuracy rate is only 90.94%. As shown in Table 1, the feature fusion classification method of the invention extracts linear combination features by combining PCA and LDA, extracts nonlinear combination features by combining PTSNE and LDA, fuses the linear combination features and the nonlinear combination features, and uses HELM as a classification model, wherein the highest classification accuracy can reach 94.881%, and the average classification accuracy can reach 93.563%. In order to verify the effect of feature fusion, the classification effect of the ELM improved algorithm serving as a classifier is compared with that of other mainstream ELM improved algorithms. The comparison result of the accuracy of the method of the invention and the mainstream electroencephalogram classification algorithm is shown in table 2, and compared with the feature extraction method combining PCA and LDA, the classification accuracy, especially the average classification accuracy of the feature fusion method is obviously improved. Compared with other comparison methods, the average classification accuracy and the highest classification accuracy of the feature fusion and classification method are the highest.

Claims

1. An EEG classification method based on HELM and combined with PTSNE and LDA feature fusion is characterized by comprising the following steps:

preprocessing an electroencephalogram signal.

Randomly disorganizing and normalizing the obtained electroencephalogram data, and segmenting the data by using overlapped sliding time windows to obtain segmented subdata;

step (2) feature extraction and fusion

a. Linear feature extraction

First-pass linear combination feature acquisition by PCA and LDA

1) The first path firstly carries out PCA dimension reduction on each segment of segmented data respectively;

2) secondly, performing LDA conversion on each segment of segmented data subjected to PCA dimensionality reduction respectively;

3) obtaining linear combination characteristics after PCA and LDA;

b. nonlinear feature extraction

The other way obtains linear combination characteristics through PCA and LDA

1) The other path firstly carries out PTSNE dimension reduction on each segment of segmented data;

2) secondly, performing LDA conversion on each segment of segmented data subjected to PTSNE dimensionality reduction respectively;

3) obtaining the nonlinear combination characteristics after PCA and LDA;

c. two-way feature fusion

Performing feature fusion on the obtained linear combination features and nonlinear combination features to serve as finally extracted features;

step (3) feature classification

And performing characteristic classification by using an HELM algorithm as a classification model of the electroencephalogram signals.

2. The electroencephalogram classification method based on HELM and combining PTSNE and LDA feature fusion as claimed in claim 1, wherein the step 1 specifically comprises the following steps: the dominant electrodes a1 and a2 were chosen, each having 896 dimensions, and the data for each electrode was divided into 9 segments, each having 128 dimensions, by a 500ms time window and 125ms overlapping window.

3. The electroencephalogram classification method based on HELM and combining PTSNE and LDA feature fusion according to claim 1, wherein the step 2 specifically comprises the following steps: the data segment divided in the A1 and A2 electrodes is copied into two equal parts, each part is respectively extracted with characteristics by different methods,

a. linear feature extraction

First-pass linear combination feature acquisition by PCA and LDA

1) The first path firstly carries out PCA dimension reduction on each segment of segmented data respectively

Obtaining the weight W by optimizing an objective function L (W)

L(w)＝w^TCw-λ(w^Tw-1)

Where C is the covariance of the input sub-data segment X,

obtaining Y from the subdata X and the obtained W

Y＝XW

Wherein,

carrying out PCA on each subdata in 18 data segments of the first data and reducing the data to D which is 16-dimensional;

2) secondly, performing LDA conversion on each segment of segmented data after PCA dimension reduction

Obtaining the weight W by optimizing an objective function L (W)

L(w)＝w^TS_Bw-λ(w^TS_Ww-1)

Wherein S is_WAnd S_BThe intra-class and inter-class covariances of the segmented sub-feature X after PCA, respectively.

Y＝W^TX

Performing LDA operation on each subdata of 18 data segments of 561 samples, respectively, each 16-dimensional feature subdata is mapped to a 1-dimensional space,

3) linear binding characteristics after PCA and LDA

Each sample has 18 1-dimensional feature subdata, the 18 1-dimensional feature subdata are combined into an 18-dimensional feature vector, and the 18-dimensional feature vector is a linear combination feature;

b. nonlinear feature extraction

1) The other path firstly carries out PTSNE dimension reduction on each segment of segmented data respectively

The mapping between the gaussian distribution in the high dimensional space and the student T distribution in the low dimensional space is learned by minimizing the KL divergence through a neural network,

the KL divergence objective function is as follows:

wherein p is_j|iIs the probability of sample j under sample i. X and Y are the high-dimensional input data and low-dimensional output characteristics, σ, of the segmented sub-data, respectively²Is the variance of the received signal and the received signal,

firstly, PTSNE operation is carried out on each subdata in 18 data segments of the second data, after PTSNE extraction characteristics, each sample has 18 data segments, each data segment has 30-dimensional characteristics,

2) secondly, performing LDA conversion on each segment of segmented data subjected to PTSNE dimension reduction respectively

Performing LDA operation on each subdata of 18 data segments in 561 samples respectively, and mapping each 16-dimensional characteristic subdata to a 1-dimensional space;

3) obtaining the nonlinear combination characteristics after PCA and LDA

Each sample has 18 1-dimensional feature sub-data, the 18 1-dimensional feature sub-data are combined into an 18-dimensional feature vector, the 18-dimensional feature vector is a nonlinear combined feature,

c. two-way feature fusion

And the first path generates linear combination features and the second path generates nonlinear combination features, and the two paths of combination features are fused together and used as final features of feature extraction and input features of the HELM.

4. The electroencephalogram classification method based on HELM and combined PTSNE and LDA feature fusion as claimed in claim 3, wherein the step 3 specifically comprises: the objective function of HELM is as follows:

β＝argmin{||Hβ-X||²+||β||_l1}

where β is the hidden layer weight, l1 is the first order normal constraint, and H is the output of the hidden layer and is defined as follows:

wherein A is_iAnd b_iThe weight matrix and the bias which are respectively initialized randomly, k is the number of the ELM layers combined with sparse self-coding, and the weight of the last layer of the HELM, namely the hidden layer of the basic ELM is as follows:

wherein C is a regularization coefficient, and I is a real label value of the identity matrix T;

for the HELM class model, the regularization coefficient C ranges from 10^-10，10^-9，…，10⁹，10¹⁰The first and second hidden layer nodes L1 are arranged in a range of 10, 20, …, 500, and the third hidden layer node L2 is arranged in a range of 100, 200, …, 1000.