KR20120122617A - Electroencephalography Classification Method for Movement Imagination and Apparatus Thereof - Google Patents

Abstract

PURPOSE: A classifying method of a brainwave imagining motion and a device thereof are provided to use support vector machine algorithm and Gaussian mixture model algorithm, thereby increasing classifying accuracy of left side and right side motion imagination electroencephalogram. CONSTITUTION: A preprocessing unit(110) receives and preprocesses measured electroencephalogram. A feature extracting unit(120) extracts a feature of the electroencephalogram. A classifying unit(130) classifies the electroencephalogram into a left side and right side motion imagination electroencephalogram class by using a support vector and support vector machine algorithm. The support vector is obtained by applying the feature extracted from left side and right side motion imagination electroencephalogram to a Gaussian mixture model. [Reference numerals] (110) Preprocessing unit; (120) Feature extracting unit; (130) Classifying unit

Description

Electroencephalography Classification Method for Movement Imagination and Apparatus Thereof}

The present invention relates to an EEG classification method and apparatus for imagining motion, and more particularly, to an EEG classification method for imagining a motion for accurately classifying a left motion imaginary EEG signal and a right motion imaginary EEG signal from measured EEG and its apparatus Relates to a device.

BCI (Brain Computer Interface) refers to an interface technology that controls a computer through brain waves by directly connecting the human brain and a computer, and BCI technology is broadly a human computer interface (HCI) technology. BCI technology is also called the Brain Machine Interface (BMI) because it allows you to manipulate machines such as wheelchairs and robots through EEG.

BCI technology has been widely used for medical purposes, and the weight of the measuring device

It was heavy and had a lot of sensors, so it was cumbersome to wear, but recently, a lightweight, easy-to-wear device in the form of a headset has been developed, and has been used for various purposes such as games and concentration improvement exercises. If BCI research and technology development are accelerated, it will be used as a next-generation interface connecting touch screen and augmented reality in the future. In particular, the classical methods such as computers, mice, and keypads used as input interfaces of computers and smartphones have recently evolved to touch pads and motion recognition, and if the development of BCI technology is accelerated, it is highly likely to be used as a next-generation interface. In particular, BCI is able to give orders without using hands or other bodies, which makes it suitable for virtual reality, image and photo recognition.

The implementation of the BCI technology receives the brain wave through a device that recognizes the brain wave stimulus, and then passes the signal processing to analyze the brain wave and give a command to the input / output device.

EEG, a measurable spontaneous electrical activity in the human scalp, is a means of understanding temporal and spatial changes in brain activity. Since it was first recorded by Hans Berger in 1929, it has been widely used in clinical and brain function research. .

EEG has both advantages and disadvantages as a computer interface. Its advantages are lower measurement cost than large devices such as fMRI, and harmless to non-invasive method where sensor is not directly applied to the scalp. In addition, it is possible to analyze the information in the brain in real time.

However, since the mixing of noise is inevitable in the EEG measurement process and loss of information is difficult to analyze, there is a problem in that EEG cannot be accurately classified.

An object of the present invention is to provide an EEG classification method and apparatus for imagining a motion for accurately classifying a left motion imaginary EEG signal and a right motion imaginary EEG signal from measured EEG.

EEG classification method for imagining the motion according to an embodiment of the present invention, the step of receiving a left motion imaginary brain wave signal and a right motion imaginary brain wave signal as training data and performing a preprocessing process, the left motion imaginary brain wave signal and the right motion imagination Extracting a feature of the EEG signal and applying the extracted feature to a Gaussian mixture model (GMM) to generate a support vector, receiving an EEG signal as test data and performing a preprocessing process, and extracting the feature of the EEG signal And classifying the EEG signal into a left motion imagined EEG signal class and a right motion imagined EEG signal class using the support vector and the SVM classification algorithm.

), 평균 벡터(

) 및 공분산 행렬(Σ _i )를 포함하며, 상기 서포트 벡터는 상기 가우시안 확률 밀도 함수의 평균 벡터(

)를 이용하여 획득될 수 있다. The Gaussian mixture model (GMM) is a combined weight of the Gaussian probability density function (

), Mean vector (

) And a covariance matrix Σ _i , wherein the support vector is an average vector of the Gaussian probability density function (

Can be obtained using

The preprocessing may include detecting an event-related synchronization (ERS) and an event-related desynchronization (ERD) from the EEG signal.

In the extracting of the statistical features of the EEG signal, a wavelet transform may be used to decompose the EEG signal and extract the statistical features of the wavelet coefficients.

The SVM algorithm can be represented by the following equation.

는 서포트 벡터, b는 바이어스 상수이며, k(?,?)는 데이터를 고차원으로 사상시키는 커널 함수이고,

는 라그랑제 승수(Lagrangian multiplier)로

이고

이다. Where t _i is -1 or 1, indicating that it corresponds to a left motion imagined EEG signal class or a right motion imagined EEG signal class,

Is a support vector, b is a bias constant, k (?,?) Is a kernel function that maps data to higher dimensions,

Is the Lagrangian multiplier

ego

to be.

In the classifying step, the EEG signal may be classified using the decision boundary obtained through the support vector.

Where y is a feature value of the EEG signal and f ( y ) is an optimized decision boundary function. If f ( y ) is greater than or equal to 0, the EEG signal corresponds to the left motion imagined EEG signal class, and if f ( y ) is less than 0, the EEG signal corresponds to the right motion imagined EEG signal class.

In accordance with another embodiment of the present invention, an EEG classification apparatus includes a preprocessor configured to receive a measured EEG signal and perform a preprocessing process, a feature extractor to extract features of the EEG signal, and a support vector and an SVM algorithm. A classification unit classifying the EEG signal into a left motion imaginary EEG signal class and a right motion imaginary EEG signal class, wherein the support vector comprises a Gaussian feature extracted from a left motion imaginary EEG signal and a right motion imaginary EEG signal. Obtained by applying to a mixed model (GMM).

According to the present invention, the classification accuracy of the left motion imaginary brain wave signal and the right motion imaginary brain wave signal can be improved by using the support vector generation algorithm (GMM) and the classification algorithm (SVM).

1 is a block diagram of an EEG classification apparatus according to an embodiment of the present invention.
2 is a flowchart illustrating a method of classifying brain waves according to an embodiment of the present invention.
3 is a characteristic distribution diagram when three elements of the GMM are used for the EEG signal.
4 is a diagram illustrating a classification result of EEG signals classified according to an EEG classification method according to an exemplary embodiment of the present invention.
5 is a view showing a result of measuring the classification accuracy according to the number of elements in the embodiment of the present invention.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1 is a block diagram of an EEG classification apparatus according to an embodiment of the present invention. As shown in FIG. 1, the EEG classification apparatus 100 according to an exemplary embodiment of the present invention includes a preprocessor 110, a feature extractor 120, and a classifier 130.

First, electroencephalography (EEG) refers to the measurement of signals generated in the brain with electrodes, and the brain waves are obtained by measuring signals on the microscopic brain surface in which electrical signals generated from numerous nerves of the brain are synthesized. As the EEG signal changes in time and space according to the activity of the brain, the state at the time of measurement, and the brain function, the EEG signal has the characteristics of each band according to frequency, characteristics in the time domain, and spatial characteristics related to brain function.

The preprocessor 110 receives the measured EEG signal, removes noise, and separates a signal for analysis. Here, EEG is multivariate time series data, and electrical signals generated in the cortex of the brain are measured through electrodes attached to a designated position of the scalp. Information useful for classification of EEG is mainly expressed in frequency component and correlates with brain function according to the location of electrode attachment.

In addition, the preprocessor 110 detects an event-related synchronization (ERS) and an event-related desynchronization (ERD) from an EEG signal to extract a feature.

ERS and ERD are important features when imagining movement in important phenomena used in the classification and analysis of EEG. ERD is a phenomenon in which the mu (μ) wave (8 to 12 Hz) is temporarily attenuated in the sensorymotor cortex when preparing or imagining movement, and ERS is a beta (β) wave temporarily after movement. 18 to 25 Hz) increases.

Motor Imagery refers to the conscious processing of movements that appear only in the thought of moving without external action, representing the psychological imagination in the area of the brain that occurs when preparing for movement. Motor Imagery can activate the primary motor area and find ERD and ERS while imagining the movement of the left or right hand.

Next, the feature extractor 120 decomposes the brain waves using a wavelet transform and extracts statistical features of the wavelet coefficients. The feature extractor 120 generates a support vector using the GMM to increase the accuracy of the classification.

The classification unit 130 classifies which class of the EEG signal belongs to the left-handed imaginary EEG signal and the right-handed imaginary EEG signal using the support vector and the SVM algorithm.

Hereinafter, a brain wave classification method according to an embodiment of the present invention will be described with reference to FIGS. 2 to 4. 2 is a flowchart illustrating a method of classifying brain waves according to an embodiment of the present invention.

First, the EEG classification apparatus 100 receives a left motion imagined brain wave signal and a right motion imagined brain wave signal as training data (S210). Then, the preprocessor 110 performs a preprocessing process on the received EEG signals to extract the features of the left motion imaginary EEG signal and the right motion imaginary EEG signal (S220). Preprocessing involves frequency analysis and ERD / ERS detection to determine the frequency bands and time zones that are characteristic of noisy brain waves.

EEG is prone to noise due to the location and method of attaching the electrode during the measurement, the experimental environment, and the movement of the subject. The EEG containing the noise is not suitable for use in a BCI system. Therefore, in order to construct an efficient BCI system, it is necessary to collect reliable EEG by performing pretreatment process. For this, autocorrelation function, independent component analysis (ICA), and band-pass filtering are performed. , Notch Filtering, Ensemble Averaging.

Next, the feature extractor 120 extracts features of the left motion imagined EEG signal and the right motion imagined EEG signal (S230). Feature extraction is the process of finding features of an input signal that have less dimensions than a given input but at the same time contain enough features to classify the data. This feature not only reduces the amount of computation for classification but also improves the classification performance by eliminating unnecessary features. In the feature extraction process, the feature extractor 120 decomposes the EEG using a wavelet transform and extracts statistical features of the wavelet coefficients.

Next, the feature extractor 120 generates a support vector using a Gaussian Mixture Model (GMM) (S240).

Hereinafter, a method of generating a support vector from a Gaussian Mixture Model (hereinafter referred to as "GMM") will be described in detail. GMM is a generation model for clustering or density estimation using statistical methods. GMM uses the EM algorithm for optimization by linearly combining the statistical distribution of several weighted Gaussian Probability Density Functions (PDFs). The GMM is expressed by M Gaussian probability density functions (PDF) as shown in Equation 1 below.

는 i 번째 PDF(또는 요소(component))를 나타내는 확률 변수이고,

는 i 번째 요소의 결합 가중치로,

로 표시하기도 한다. 그리고 확률 값이므로

,

을 만족한다.

는 n 차원 벡터

의 평균 벡터

와

대각행렬의 공분산 Σ _i 를 나타낸다. 여기서

는 추출된 특성에 대한 특성 벡터를 나타낸다.

은 파라미터

에 의한 i 번째 요소의 가우시안 확률 밀도 함수를 나타낸다. 가우시안 확률 밀도 함수(

)는 다음의 수학식 2와 같이 표현된다.here

Is a random variable representing the i th PDF (or component),

Is the combined weight of the i th element,

It is also indicated by. And the probability value

,

To satisfy.

N dimension vector

Mean vector of

Wow

Covariance Σ _i of the diagonal matrix is shown. here

Denotes a feature vector for the extracted feature.

Is a parameter

Represents a Gaussian probability density function of the i th element by. Gaussian Probability Density Function (

) Is expressed by Equation 2 below.

), 평균 벡터(

) 그리고 공분산 행렬(Σ _i )을 이용해 수학식 3과 같이 표현할 수 있다. 이때 최적 모델을 추정하기 위해 MLE(maximum likelihood estimate)를 사용하고 이를 위한 파라미터는 EM 알고리즘을 반복적으로 학습하여 구한다. EM 알고리즘으로 생성된 수학식 3과 같은 고차원 벡터를 GMM-UBM (universal background model)이라고 하고 수학식 3과 같이 표현한다.And GMM is the join weight (

), Mean vector (

) And the covariance matrix Σ _i can be expressed as Equation 3. At this time, the maximum likelihood estimate (MLE) is used to estimate the optimal model, and the parameters for this are obtained by repeatedly learning the EM algorithm. A high-dimensional vector such as Equation 3 generated by the EM algorithm is called GMM-UBM (universal background model) and is expressed as Equation 3.

만을 사용하여 클래스의 결정 경계를 결정하는 서포트 벡터를 생성한다.The feature extractor 120 is an average vector of parameters of an optimal model generated through MLE.

Use only to create a support vector that determines the class's decision boundary.

), 평균 벡터(

) 및 공분산 행렬(Σ _i ) 중에서 가우시안 확률 밀도 함수, 즉 요소의 평균 벡터(

)만을 이용하여 서포트 벡터를 생성한다. That is, the feature extractor 120 combines the combined weights constituting the GMM (

) , Mean vector (

) And the covariance matrix Σ _i , a Gaussian probability density function, that is, the mean vector of elements (

) To create a support vector.

)를 3개로 정했을 경우 각각의 클래스를 나타내는 특징 분포에 대해 3개의 요소가 생성된 결과를 나타낸다. 3 is a characteristic distribution diagram when three elements of the GMM are used for the EEG signal. That is, FIG. 3 extracts the features of the left motion imaginary brain wave (indicated in blue) and the right motion imaginary brain wave (indicated in red) for the motion imaginary brain wave, respectively,

If you specify 3), it shows the result of generating 3 elements for the feature distribution representing each class.

As described above, according to the exemplary embodiment of the present invention, a support vector is generated by using training data for the left motion imaginary brain wave and the right motion imaginary brain wave. Hereinafter, a method of classifying a left-motion imaginary brain wave and a right-motion imaginary brain wave using the actually measured brain wave signal will be described. Duplicate descriptions will be omitted in the EEG classification process.

First, the EEG classification apparatus 100 receives an EEG signal from the outside as test data (Test Data) (S250). In addition, the preprocessor 110 performs a preprocessing process on the received EEG signal (S260). That is, the preprocessor 110 detects the frequency analysis and the ERD / ERS from the received EEG signal in order to know the frequency band and time zone which are characteristic of the noisy EEG.

Next, the feature extractor 120 extracts a feature from the preprocessed EEG signal (S270). That is, the feature extractor 120 decomposes an EEG signal using a wavelet transform and extracts statistical features of wavelet coefficients.

In addition, the classification unit 130 classifies the brain waves into the left motion imaginary brain wave and the right motion imaginary brain wave using the support vector generated in step S240 (S280). The classifier 130 classifies the EEG signals using a support vector machine (SVM) classification algorithm.

Here, SVM, a classification algorithm, is an identification model widely used for many linear and nonlinear classification problems because of its high prediction accuracy, and is an algorithm for minimizing structural risk.

The classifier 130 finds a crystal hyperplane that applies a support vector to the SVM algorithm to maximize the margin between two classes (imaginary EEG). The obtained hyperplane is also called a crystal boundary, and the data of the classes closest to the crystal boundary becomes the support vector.

Therefore, the classifying unit 130 classifies the received EEG signal into the left motion imaginary brain wave and the right motion imaginary brain wave signal on the basis of the crystal hyperplane. That is, when classifying a nonlinear signal such as an EEG, the classifying unit 130 performs linear separation by mapping data in a high dimension using a kernel. The SVM classification algorithm is defined as in Equation 4 below.

는 서포트 벡터, b는 2차 분류 문제를 풀기 위한 바이어스 상수이다. k(?,?)는 데이터를 고차원으로 사상시키는 커널 함수이다.

는 조건부 최적화 문제를 해결하기 위해 사용하는 라그랑제 승수(Lagrangian multiplier)로

이고

이다. Where t _i is an ideal output of -1 or 1, indicating that it corresponds to class 0 or 1

Is a support vector, and b is a bias constant for solving the quadratic classification problem. k (?,?) is a kernel function that maps data to higher dimensions.

Is the Lagrangian multiplier used to solve the conditional optimization problem.

ego

to be.

The classifier 130 classifies the test data, that is, the received EEG signal, as shown in Equation 5 using the optimized decision boundary obtained through the support vector.

Where y is the feature value of the test data and f ( y ) is the optimized decision boundary function. If f ( y ) is greater than or equal to 0, the test data corresponds to class 1; if f ( y ) is less than 0, the test data corresponds to class -1.

As such, the classifying unit 130 classifies the class by determining an optimized decision boundary between the classes (left or right motion imagined brain waves) using the SVM.

4 is a diagram illustrating a classification result of EEG signals classified according to an EEG classification method according to an exemplary embodiment of the present invention. That is, Figure 4 shows the results of classifying the EEG using the support vector and the SVM classification algorithm generated by the GMM.

FIG. 5 is a diagram illustrating a result of measuring classification accuracy according to the number of elements in the embodiment of the present invention. Table 1 shows the accuracy when classifying EEG using only the SVM classification algorithm and the GMM as shown in FIG. 5. The results show the accuracy of classifying EEG using the SVM classification algorithm.

As shown in FIG. 5 and Table 1, when only the SVM is applied to the motion imaginary EEG signal, a maximum of 81.63% of the result is obtained. In this case, a maximum of 83.61% resulted in improved classification accuracy.

As described above, according to the exemplary embodiment of the present invention, the left motion imaginary brain wave signal and the right motion imaginary brain wave signal can be accurately classified using the support vector generation algorithm (GMM) and the classification algorithm (SVM).

While the present invention has been described with reference to exemplary embodiments, it is to be understood that the invention is not limited to the disclosed embodiments, but, on the contrary, is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims. Accordingly, the true scope of the present invention should be determined by the technical idea of the appended claims.

100: EEG classification apparatus, 110: preprocessing unit,
120: feature extraction section, 130: classification section

Claims (12)

Receiving a left motion imagined brain wave signal and a right motion imagined brain wave signal as training data and performing a preprocessing process,
Extracting features of the left motion imagined EEG signal and right motion imagined EEG signal, and applying the extracted features to a Gaussian mixture model (GMM) to generate a support vector;
Receiving an EEG signal as test data and performing a preprocessing process,
Extracting features of the EEG signal, and
And classifying the EEG signal into a left motion imagined EEG signal class and a right motion imagined EEG signal class using the support vector and the SVM classification algorithm. The method of claim 1,
The Gaussian mixture model (GMM) is a combined weight of the Gaussian probability density function (

), Mean vector (

) And the covariance matrix (Σ _i ),
The support vector is an average vector of the Gaussian probability density function (

EEG classification method imagining the movement obtained using The method of claim 2,
The pre-
EEG classification method for imagining the movement comprising the step of detecting the event-related synchronization (ERS) and the event-related desynchronization (ERD) from the EEG signal. The method of claim 2,
Extracting a statistical feature of the EEG signal,
EEG classification method using a wavelet transform to imagine the motion to decompose the EEG signal and extract the statistical characteristics of the wavelet coefficients. The method of claim 2,
The SVM algorithm is a brain wave classification method for imagining a motion represented by the following equation:

Where t _i is -1 or 1, indicating that it corresponds to a left motion imagined EEG signal class or a right motion imagined EEG signal class,

Is a support vector, b is a bias constant, k (?,?) Is a kernel function that maps data to higher dimensions,

Is the Lagrangian multiplier

ego

to be. The method of claim 5,
The classifying step,
EEG classification method for imagining the motion to classify the EEG signal using the decision boundary obtained through the support vector:

Where y is a feature value of the EEG signal and f ( y ) is an optimized decision boundary function. If f ( y ) is greater than or equal to 0, the EEG signal corresponds to the left motion imagined EEG signal class, and if f ( y ) is less than 0, the EEG signal corresponds to the right motion imagined EEG signal class. A preprocessor for receiving the measured EEG signal and performing a preprocessing process,
Feature extraction unit for extracting the features of the EEG signal, And
A classification unit for classifying the EEG signal into a left motion imagined EEG signal class and a right motion imagined EEG signal class using a support vector and an SVM algorithm,
The support vector is
The training data is obtained by applying a feature extracted from a left motion imagined EEG signal and a right motion imagined EEG signal to a Gaussian mixture model (GMM).
EEG classifier to imagine movement. The method of claim 7, wherein
The Gaussian mixture model (GMM) is a combined weight of the Gaussian probability density function (

), Mean vector (

) And the covariance matrix (Σ _i ),
The support vector is an average vector of the Gaussian probability density function (

EEG classification device for imagining the movement obtained by using. 9. The method of claim 8,
The preprocessing unit,
An EEG classification method imagining a movement for detecting ERS (Event-Related Synchronization) and ERD (Event-Related Desynchronization) from the EEG signal. 9. The method of claim 8,
The feature extraction unit,
EEG classification apparatus using a wavelet transform to imagine the motion of decomposing the EEG signal and extracting the statistical features of the wavelet coefficients. 9. The method of claim 8,
The SVM algorithm is an EEG classifier that imagines a motion represented by the following equation:

Where t _i is -1 or 1, indicating that it corresponds to a left motion imagined EEG signal class or a right motion imagined EEG signal class,

Is a support vector, b is a bias constant, k (?,?) Is a kernel function that maps data to higher dimensions,

Is the Lagrangian multiplier

ego

to be. The method of claim 11,
Wherein,
EEG classification apparatus for imagining the motion to classify the EEG signal using the crystal boundary obtained through the support vector:

Where y is a feature value of the EEG signal and f ( y ) is an optimized decision boundary function. If f ( y ) is greater than or equal to 0, the EEG signal corresponds to the left motion imagined EEG signal class, and if f ( y ) is less than 0, the EEG signal corresponds to the right motion imagined EEG signal class.

Images