CN106503616A - A kind of Mental imagery Method of EEG signals classification of the learning machine that transfinited based on layering - Google Patents

Abstract

The invention discloses a motor imagery EEG signal classification method based on a layered extreme learning machine, which belongs to the field of pattern recognition and brain-computer interface. First, the original signal of each sample is segmented by windowing to obtain S-segment sub-signals; then, principal component analysis and linear discriminant analysis are performed on the S-segment sub-signals, and the final S K-1-dimensional feature vectors are processed Combined to obtain S*(K-1)-dimensional features; finally, the features of all samples are sent to the hierarchical extreme learning machine classifier to obtain the final classification result. The traditional ELM algorithm is a single hidden layer structure, which has great limitations in analyzing complex signals. The present invention uses HELM for classification and recognition, and uses the sparse self-encoding based on ELM to transform the single hidden layer ELM into a deep network structure. Hierarchical feature extraction and classification extracts deep information, improves classification accuracy, and maintains the advantage of low time-consuming ELM.

Description

A Classification Method of Motor Imagery EEG Signals Based on Hierarchical Extreme Learning Machine

technical field

The invention belongs to the fields of pattern recognition and Brain-Computer Interface (Brain-Computer Interface, BCI), and relates to a method for classifying motor imagery EEG signals in a Brain-Computer Interface system device by using a layered extreme learning machine.

Background technique

In today's society, the problem of population aging is prominent. With the continuous advancement of science and technology, the research of artificial intelligence in medical treatment is deepening. In particular, the application and development of brain-computer interface technology has far-reaching significance for the diagnosis, treatment and rehabilitation of diseases. Brain-computer interface technology builds a communication bridge between the human brain and external devices, allowing the human brain to control external devices while obtaining information from the human brain. Motor imagery EEG signals are widely used in the field of brain-computer interface. By analyzing the EEG signals of the human brain during imaginative activities, the state of the brain can be identified, and then the purpose of controlling external devices can be achieved. big help.

Brain-computer interface technology can be divided into five parts, namely: signal extraction, signal preprocessing, feature extraction, feature classification, and interface control. EEG signal classification and recognition is one of the key technologies, mainly including feature extraction and feature classification. EEG signals pose great challenges to classification and recognition due to their complex characteristics such as high dimensionality, randomness, and instability. Therefore, finding effective feature extraction and classification methods is the key to improving the accuracy of EEG signal recognition. At the same time, the practical application of brain-computer interface puts forward requirements on the timeliness of equipment response. Therefore, classification accuracy and time consumption are important indicators for evaluating the classification and recognition system of EEG signals.

The Extreme Learning Machine (ELM) method based on Single-hidden Layer Feedforward Neural Network (SLFN) proposed by Huang, its training speed is comparable to that of BP neural network and support vector machine (SVM) Significantly improved compared to . On this basis, the sparse autoencoder based on ELM is introduced for multi-level feature extraction, and the final high-level features are classified by ELM. This structure is called Hierarchical Extreme Learning Machine (HELM). On the basis of maintaining the advantages of fast ELM training speed, HELM expands it into a deep structure, increases the ability to process complex signals, and improves classification performance.

Contents of the invention

In view of the above problems, the present invention adopts a classification method based on HELM to classify the EEG signals of the motor imagery task, so as to improve the classification accuracy and classification speed.

The main train of thought of realizing the method of the present invention is: at first, carry out windowing segmentation to the original signal of each sample, obtain S-segment sub-signal; Then, carry out principal component analysis and linear discriminant analysis to this S-segment sub-signal respectively, and final S K-1 dimensional feature vectors are combined to obtain S*(K-1) dimensional features; finally, the features of all samples are sent to the hierarchical ELM classifier to obtain the final classification result.

The motor imagery EEG signal classification method based on layered extreme learning machine comprises the following steps:

Step 1, using a fixed sliding time window to divide the original motor imagery EEG signal into S-segment sub-signals. The value of S depends on the length of the sliding time window and the length of the original EEG signal.

In step 2, the principal component analysis (PCA) method is used to reduce the dimensionality of each sub-signal obtained in step 1 to reduce redundant information in the signal, that is, to use principal components to represent signals with similar characteristics, and finally obtain the reduced dimension eigenvectors of .

In step 3, the feature vector obtained in step 2 is subjected to secondary dimensionality reduction through a linear discriminant analysis (Liner Discriminate Analysis, LDA) method, and K-1 dimensional feature vectors are obtained for K categories of EEG data. For binary classification problems, a one-dimensional feature vector is obtained.

Step 4, each sub-signal is processed according to step 2 and step 3. For S segment sub-signals, S K-1-dimensional feature vectors will be obtained in the end, and S K-1-dimensional feature vectors are combined to obtain The final feature is S*(K-1) dimension.

Step five, send the S*(K-1) dimensional features obtained in step four to the HELM classifier to obtain the final classification result.

Compared with the prior art, the method of the present invention has the following advantages:

The traditional ELM algorithm is a single hidden layer structure, which has great limitations in analyzing complex signals. The present invention uses HELM for classification and recognition, and uses the sparse self-encoding based on ELM to transform the single hidden layer ELM into a deep network structure. Hierarchical feature extraction and classification extracts deep information, improves classification accuracy, and maintains the advantage of low time-consuming ELM.

Description of drawings

Fig. 1 is the main flowchart of the method involved in the present invention;

Fig. 2 is a schematic diagram of the HELM classification method involved in the present invention.

detailed description

The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

Suppose there is a training data set TrainData and a set of test data sets TestData, the sample size of TrainData is N, and the dimension is D; the sample size of TestData is M, and the dimension is also D. Among them, the samples in TrainData and TestData belong to K categories.

The flow chart of the motor imagery EEG signal classification method based on hierarchical extreme learning machine is shown in Figure 1.

Step 1: Divide the TrainData and TestData into S-segment EEG signals by means of fixed time window division. TrainData _i represents the i-th sub-signal in the training data set, and the dimension of each sub-signal is D _i (i=1,2,...,S). TestData _i represents the i-th sub-signal in the test data set, and the dimension of each sub-signal is D _i (i=1,2,...,S). Because the fixed time window is adopted, the size of the window is fixed, so D ₁ =D ₂ =...=D _i .

There are two types of fixed sliding time windows: one is non-overlapping time windows, and there is no overlap between each sub-signal, namely The other is a superimposed time window, and there is a partial overlap between every two adjacent sub-signals, that is

Step 2: Perform dimensionality reduction on each segment of the sub-signals TrainData _i and TestData _i obtained in Step 1 by principal component analysis. After sorting the eigenvalues from large to small, according to the cumulative contribution rate, only retain the eigenvector combination M _PCA corresponding to the first m largest eigenvalues as the projection space vector, M _PCA = [Φ ₁ ,Φ ₂ ,... ,Φ _m ], Φ ₁ ,Φ ₂ ,...,Φ _m are respectively the eigenvectors corresponding to the largest eigenvalues obtained after dimensionality reduction by principal component analysis. Simultaneously project TrainData _i and TestData _i onto M _PCA to obtain training data Train _i and test data Test _i after PCA dimensionality reduction:

Train _i =TrainData _i M _PCA

Test _i =TestData _i M _PCA

In step 3, the feature vector obtained in step 2 is subjected to secondary dimensionality reduction by the LDA method, and the specific method is as follows:

(1) According to the LDA criterion, the projected space vector w ^* of LDA is calculated by using the inter-class scatter matrix of different classes of samples in Train _i and the intra-class scatter matrix of samples of the same class.

(2) Project Train _i and Test _i onto w ^* to obtain the characteristics of the i-th segment of the EEG signal:

Trainfeature _i ＝Train _i ·w ^*

Testfeature _i ＝Test _i ·w ^*

Step 4, calculate all Trainfeature _i and Testfeature _i and combine them to get the final features TrainFeature and TestFeature:

TrainFeature = { _{TrainFeature1} , _{TrainFeature2} , ..., TrainFeature _x }

TestFeature = { _TestFeature1 , _TestFeature2 , ..., TestFeature _x }

The entire feature extraction process is shown in Figure 1.

Step 5, use the feature TrainFeature obtained in step 4 to train the HELM classifier model, send TestFeature to the trained model for classification, and obtain the final classification result. The flow chart is shown in Figure 2, and the specific method is as follows:

(1) The number of hidden layers K, the number of hidden layer nodes L={L ₁ , L ₂ ,...,L _K } and the activation function g(x) are given. Randomly generate input weight a _i and bias value b _i . x _i represents the i-th training sample input, t _i corresponds to the label of the i-th training sample, H _i is the output matrix of the i-th hidden layer, and A _i is the hidden layer output matrix of the i-th sparse autoencoder. Because the extracted features are sent to the classifier, so in the present invention, _xi actually represents TrainFeature _i .

(2) Calculate the hidden layer weight β _i of the sparse autoencoder based on ELM layer by layer, i={1, 2, ..., K-1, K}, where H ₀ =X, the specific calculation steps are as follows:

a) Calculate the hidden layer output A _i of the sparse autoencoder:

A _i =g(a _i ·H _i-1 +b _i )

g(·) is the selected activation function g(x).

b) Calculation of the Lipschitz constant L _i :

L _i ＝λ _max (A _i ^T ·A _i )

λ _max is the largest eigenvalue.

c) Iteratively calculate the hidden layer weight β _i , where the number of iterations j=1:50, the initial time point t ₁ =1, and the specific value of β _i is initially y ₁ =β _i0 , the calculation process is as follows:

β _j ＝{‖c _j ‖,0} _max *sign(c _j )

β _i = β _j

d) Calculate the hidden layer output matrix H _i :

(3) Calculate the hidden layer output matrix H _K+1 of the classification part.

(4) Find the optimal weight β of the classification part:

According to the target output of the neural network T=H _K+1 β. If H _K+1 is full rank, the optimal weight is obtained by the least square method, and the solution is:

is the generalized inverse matrix of H K+ ₁ , if H _K+1 is non-full rank, then use singular value decomposition to solve the generalized inverse matrix of H _K+1 To calculate the optimal weight β.

(5) Send TestFeature into the model and calculate layer by layer to obtain a set of M-dimensional prediction labels, and compare the prediction labels with the real labels to obtain the classification accuracy.

In this embodiment, the BCI2003Ia standard data set is selected, which is two types of motor imagery EEG data, and the correct rate of classification results is up to 94.54%. The correct rate of the embodiment of the present invention is higher than that of the methods adopted by other scholars studying the data. And in the case of the same characteristics, the classification effect is better than 92.15% of SVM and 89.04% of the average result of ELM. Compared with the multi-layer extreme learning machine with the same deep network structure, the classification result is increased by 0.34%, and the time-consuming is shortened by about half. This invention is also applicable to multi-classification problems, which can be converted into two classifications first, and continue to use the method used in the present invention.

Claims (3)

1. A motor imagery EEG signal classification method based on layered extreme learning machine, characterized in that: First, the original signal of each sample is segmented by windowing to obtain S-segment sub-signals; then, principal component analysis and linear discriminant analysis are performed on the S-segment sub-signals, and the final S K-1 dimensional feature vectors are Combining to obtain S*(K-1)-dimensional features; finally, sending the features of all samples into the hierarchical extreme learning machine classifier to obtain the final classification result; The motor imagery EEG signal classification method based on layered extreme learning machine comprises the following steps: Step 1, using a fixed sliding time window to divide the original motor imagery EEG signal into S segment sub-signals; the value of S depends on the length of the sliding time window and the length of the original EEG signal; Step 2: Perform dimensionality reduction on each sub-signal obtained in step 1 by principal component analysis method to reduce redundant information in the signal, that is, use principal components to represent signals with similar characteristics, and finally obtain dimensionality-reduced feature vectors; Step 3, the eigenvector obtained in step 2 is subjected to secondary dimensionality reduction through the linear discriminant analysis method, and for K categories of EEG data, a K-1 dimensional eigenvector is obtained; for a binary classification problem, a one-dimensional eigenvector of dimension; Step 4, each sub-signal is processed according to step 2 and step 3. For S segment sub-signals, S K-1-dimensional feature vectors will be obtained in the end, and S K-1-dimensional feature vectors are combined to obtain The final feature is S*(K-1) dimension; Step five, send the S*(K-1) dimensional features obtained in step four to the HELM classifier to obtain the final classification result. 2. a kind of motor imagery EEG signal classification method based on layered extreme learning machine according to claim 1, is characterized in that: assume training data set TrainData and a group of test data set TestData, the sample size of TrainData is N, the dimension is D; the sample size of TestData is M, and the dimension is also D; the samples in TrainData and TestData belong to K categories; Motor imagery EEG signal classification method based on hierarchical extreme learning machine, Step 1: Divide TrainData and TestData into S-segment EEG signals by means of fixed time window division; TrainData _i represents the i-th sub-signal in the training data set, and the dimension of each sub-signal is D _i (i=1,2,..., S); TestData _i represents the i-th sub-signal in the test data set, and the dimension of each sub-signal is D _i (i=1,2,...,S); because the fixed time window is adopted, the window size is fixed, so D ₁ = D ₂ = . . . = D _i ; There are two types of fixed sliding time windows: one is non-overlapping time windows, and there is no overlap between each sub-signal, namely The other is a superimposed time window, and there is a partial overlap between every two adjacent sub-signals, that is Step 2: Reduce the dimensionality of each segment of the sub-signals TrainData _i and TestData _i obtained in Step 1 through principal component analysis; sort the eigenvalues from large to small, and then keep only the top m largest ones according to the cumulative contribution rate The eigenvector combination M _PCA corresponding to the eigenvalue is used as the projection space vector, M _PCA = [Φ ₁ , Φ ₂ ,..., Φ _m ], Φ ₁ , Φ ₂ ,..., Φ _m are respectively through principal component analysis The method obtains the eigenvector corresponding to the largest eigenvalue after dimensionality reduction; the _{TrainData i} and TestData _i are simultaneously projected onto MPCA to obtain the training data Train _i and test data Test _i after PCA dimensionality reduction: Train _i =TrainData _i M _PCA Test _i =TestData _i M _PCA In step 3, the feature vector obtained in step 2 is subjected to secondary dimensionality reduction by the LDA method, and the specific method is as follows: (1) According to the LDA criterion, the projected space vector w ^* of LDA is calculated by using the inter-class scatter matrix of different class samples and the intra-class scatter matrix of the same class sample in Train _i ; (2) Project Train _i and Test _i onto w ^* to obtain the characteristics of the i-th segment of the EEG signal: Trainfeature _i ＝Train _i ·w ^* Testfeature _i ＝Test _i ·w ^* Step 4, calculate all Trainfeature _i and Testfeature _i and combine them to get the final features TrainFeature and TestFeature: TrainFeature = { _{TrainFeature1} , _{TrainFeature2} , ..., TrainFeature _x } TestFeature = { _TestFeature1 , _TestFeature2 , ..., TestFeature _x } Step 5, use the feature TrainFeature obtained in step 4 to train the HELM classifier model, send TestFeature to the trained model for classification, and obtain the final classification result. 3. a kind of motor imagery EEG signal classification method based on layered extreme learning machine according to claim 2, is characterized in that: The specific method of step five is as follows: (1) Given the number of hidden layers K, the number of nodes in the hidden layer L={L ₁ ,L ₂ ,…,L _K } and the activation function g(x); randomly generate input weight a _i and bias value b _i ; x _i represents the i-th training sample input, t _i corresponds to the label of the i-th training sample, H _i is the output matrix of the i-th hidden layer, and A _i is the hidden layer output of the i-th sparse autoencoder Matrix; because it is the feature that has been extracted into the classifier, so in the present invention, x _i actually represents TrainFeature _i ; (2) Calculate the hidden layer weight β _i of the sparse autoencoder based on ELM layer by layer, i={1, 2, ..., K-1, K}, where H ₀ =X, the specific calculation steps are as follows: a) Calculate the hidden layer output A _i of the sparse autoencoder: A _i =g(a _i ·H _i-1 +b _i ) g( ) is the selected activation function g(x); b) Calculation of the Lipschitz constant L _i : L _i ＝λ _max (A _i ^T ·A _i ) λ _max is the largest eigenvalue; c) Iteratively calculate the hidden layer weight β _i , where the number of iterations j=1:50, the initial time point t ₁ =1, and the specific value of β _i is initially y ₁ =β _i0 , the calculation process is as follows: ${\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; 2</span>}}{{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}<span\; class="notranslate">\; \&CenterDot;</span><span\; class="notranslate">\; (</span>{{\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; \&Center\; Dot;</span>{\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&Center\; Dot;</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; +</span>\frac{\mathrm{<span\; class="notranslate">\; 2</span>}}{{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}<span\; class="notranslate">\; \&Center\; Dot;</span>{{\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; \&Center\; Dot;</span>{\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}$ β _j ＝{‖c _j ‖,0} _max *sign(c _j ) ${\mathrm{<span\; class="notranslate">\; t</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; +</span>\sqrt{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 4</span>}{\mathrm{<span\; class="notranslate">\; t</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}}{\mathrm{<span\; class="notranslate">\; 2</span>}}$ ${\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&beta;</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; +</span><span\; class="notranslate">\; (</span>\frac{{\mathrm{<span\; class="notranslate">\; t</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; t</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&beta;</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&beta;</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; )</span>$ β _i = β _j d) Calculate the hidden layer output matrix H _i : ${\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; g</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; \&Center\; Dot;</span>{\mathrm{<span\; class="notranslate">\; \&beta;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; )</span>$ (3) Calculate the hidden layer output matrix H _K+1 of the classification part; (4) Find the optimal weight β of the classification part: According to the target output of the neural network T=H _K+1 β; if H _K+1 is full-ranked, the optimal weight is obtained by the least square method, and the solution is: yes The generalized inverse matrix of H K+1, if H _K+1 is non-full rank, then use the singular value decomposition to solve the generalized inverse matrix of H _K+1 To calculate the optimal weight β; (5) Send TestFeature into the model and calculate layer by layer to obtain a set of M-dimensional prediction labels, and compare the prediction labels with the real labels to obtain the classification accuracy.