KR101380964B1 - Apparatus of brain-computer interface, and classification method thereof - Google Patents

Abstract

A brain-computer connection device for classifying test signals, which are brain wave signals, into any one of a plurality of classes, comprising: a dictionary unit for designing a dictionary for classifying test signals based on training signals of each class, and a solution of test signals based on a dictionary And a classifier for classifying the test signal based on the distribution of nonzero coefficients shown in each class of the solution.

Description

Brain-computer connection device and its classification method {APPARATUS OF BRAIN-COMPUTER INTERFACE, AND CLASSIFICATION METHOD THEREOF}

The present invention relates to a brain-computer connection device and a classification method thereof.

To date, studies have shown that electroencephalograms (EEGs) measured from the scalp can communicate even with little spontaneous movement. This is commonly called an EEG-based brain-computer interface (BCI) system. The BCI system extracts specific characteristic signals of EEG activity and uses them to generate control signals. Motion imagery-based BCI (BCI) systems use sensorimotor rhythms (SMRs) that include Mu (8-14 Hz) and Beta (15-30 Hz) rhythms in the sensorimotor cortex of the scalp. . When the subject imagines left or right hand movements, the activity of EEG signals such as attenuation of the Mu rhythm frequency band and event related desynchronisation (ERD) occurs in the part of the sensory cortex opposite the imaginary hand.

The BCI system translates the signal feature detected from the test subject into a command signal, which is performed through a classification algorithm. However, it is not easy to classify the motion imaginary signal used by the BCI system because it is very noisy and varies from subject to subject.

The problem to be solved by the present invention is to provide a brain-computer connection device for classifying a motion imaginary brain wave signal based on the sparse expression, and a classification method thereof.

A brain-computer connection device for classifying a test signal, which is an EEG signal, according to an embodiment of the present invention into any one of a plurality of classes, wherein a dictionary for classifying the test signal is designed based on a training signal of each class And a classification unit for classifying the test signal based on a distribution of non-zero coefficients shown in each class of the solution.

The brain-computer connection device may further include a CSP filtering unit configured to output a CSP filtered signal by performing CSP filtering to maximize the training signal for one class and minimize the other class, and the dictionary unit is configured to filter the CSP. The dictionary can be designed based on the signal.

The dictionary unit may be configured of a plurality of matrices corresponding to each of the plurality of classes, the number of rows is the number of CSP filters, and the number of columns of each class is the number of training signals of each class.

The brain-computer connection apparatus may further include a feature signal extracting unit extracting a signal of a specific frequency band from the CSP filtered signal as a feature signal, and the dictionary unit may design the dictionary based on the feature signal.

The specific frequency band may be a frequency band corresponding to the sensory rhythm of the sensorimotor cortex portion of the scalp.

The solver can obtain the solution by L1 minimization.

The classifying unit may classify the test signal into a class in which a non-zero coefficient is highest among each class of the solution.

The classification unit calculates a residual for each class based on the solution to classify the test signal into a class having a minimum residual, and the residual for each class includes a vector generated by filling an element of a class other than itself with zero in the solution. It may be a value calculated as a difference between a first value projected in advance and a second value corresponding to the test signal.

The training signal and the test signal may be a motion imagination based EEG signal.

The plurality of classes may include a left class corresponding to a left movement imagination and a right class corresponding to a right movement imagination.

A method of classifying a test signal as an EEG signal into any one of a plurality of classes by a brain-computer connection device according to another embodiment of the present invention, the method comprising: acquiring a training signal corresponding to each class, and classifying the training signal into one class Generating a CSP filtered signal by maximizing the CSP and minimizing the other classes by using the CSP filtering signal; designing a dictionary for solving the test signal based on the CSP filtered signal; Obtaining a solution of the test signal on the basis of and classifying the test signal based on a distribution of nonzero coefficients represented in each class of the solution.

The training signal and the test signal may be a motion imagination based EEG signal.

The plurality of classes may include a left class corresponding to a left movement imagination and a right class corresponding to a right movement imagination.

The designing the dictionary may include a plurality of matrices corresponding to each of the plurality of classes, the number of rows is the number of CSP filters, and the number of columns of each class is the number of training signals of each class. Can be.

The classification method may further include extracting a signal of a specific frequency band from the CSP filtered signal as a feature signal, and designing the dictionary may design the dictionary based on the feature signal.

The specific frequency band may be a frequency band corresponding to the sensory rhythm of the sensorimotor cortex portion of the scalp.

Obtaining the solution may be obtained by the L1 minimization method.

The obtaining of the solution may obtain the solution so that the number of nonzero coefficients may be determined to be less than the number of training signals of each class.

In the classifying of the test signal, the test signal may be classified into a class in which a non-zero coefficient is highest among each class of the solution.

The classifying the test signal may include classifying the test signal into a class having a minimum residual, by calculating a residual for each class based on the solution, and filling the element of a class other than itself among the solutions with zeros. The generated vector may be a value calculated by a difference between a first value projected beforehand and a second value corresponding to the test signal.

According to an exemplary embodiment of the present invention, a motion imaginary EEG signal which is noisy and changes over time can be classified more accurately than a conventional classification method. In particular, when classifying a test signal that is difficult to discriminate, it may be classified more accurately than a conventional linear discrimination method.

1 is a schematic block diagram of a brain-computer connection device according to an embodiment of the present invention.
2 is a schematic block diagram of a brain-computer connection device according to another embodiment of the present invention.
3 is a diagram illustrating a dictionary designed according to an embodiment of the present invention.
4 is a diagram illustrating a linear equation according to an embodiment of the present invention.
5 is a diagram illustrating a training signal relationship before CSP filtering according to an embodiment of the present invention.
6 is a diagram illustrating a training signal relationship after CSP filtering according to an embodiment of the present invention.
7 is a flowchart of a classification method according to an embodiment of the present invention.
8 is a diagram showing an easy classification result by the linear discriminant method.
9 is a diagram illustrating a classification result according to an embodiment of the present invention.
10 is a diagram showing a difficult classification result by the linear discriminant method.
11 is a view showing a classification result according to another embodiment of the present invention.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art can easily carry out the present invention. The present invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. In order to clearly illustrate the present invention, parts not related to the description are omitted, and similar parts are denoted by like reference characters throughout the specification.

Throughout the specification, when an element is referred to as "comprising ", it means that it can include other elements as well, without excluding other elements unless specifically stated otherwise.

A brain-computer connection device and a classification method thereof according to an embodiment of the present invention will now be described with reference to the drawings.

1 is a schematic block diagram of a brain-computer connection device according to an embodiment of the present invention.

)를 어느 하나의 클래스(class)로 분류한다. 여기서, 클래스는 테스트 신호(

)를 분류하기 위해 나누어진 단위로서, 복수 개일 수 있다. 간단히, 클래스는 왼쪽 클래스와 오른쪽 클래스로 가정하고 설명한다. 여기서 희소 표현(sparse representation)이란 다양한 자연신호가 특정 기저에서는 희소(sparse)하기 때문에, 압축센싱(compressive sensing, CS)을 통해 자연신호들을 희소하게 표현(Sparse Representation)할 수 있다는 이론이다. 그리고, 이렇게 희소하게 표현 가능한 신호의 각 샘플은 모든 신호의 전체적인 관점을 담고 있기 때문에 간단한 선형적 투영 연산을 통하여 홀리스틱(holistic) 샘플로 압축될 수 있다. 이런 홀리스틱 샘플들은 양호한 신호 복구에 필요한 압축된 샘플들의 개수가 섀넌-나이퀴스트(Shannon-Nyquist) 샘플 개수보다 상당히 적다.Referring to FIG. 1, the brain-computer connection apparatus 100 may test a test signal through a sparse representation-based classification (SRC).

) Into one class. Where class represents the test signal (

) Are divided into units, which may be plural. For simplicity, we assume that classes are left and right classes. Here, the sparse representation is a theory that sparse representation of natural signals can be sparse through compression sensing (CS) because various natural signals are sparse at a specific basis. Each sample of such a sparse signal can be compressed into a holistic sample through a simple linear projection operation because it contains an overall view of all signals. These holistic samples have a significantly smaller number of compressed samples for good signal recovery than Shannon-Nyquist samples.

)를 사전(dictionary,

)의 희소 표현을 기초로 해(

)를 구한다. 그리고 뇌-컴퓨터 접속 장치(100)는 해(

)를 기초로 테스트 신호(

)를 분류한다. 이를 위해, 뇌-컴퓨터 접속 장치(100)는 사전부(dictionary unit)(110), 풀이부(solving unit)(130), 그리고 분류부(classification unit)(150)를 포함한다. The brain-computer connection device 100 is a signal to be classified, that is, a test signal (

) Into a dictionary,

Based on the rare expression of)

). And brain-computer connection device 100

Based on the test signal (

Classify To this end, the brain-computer connection device 100 includes a dictionary unit 110, a solving unit 130, and a classification unit 150.

)을 설계한다. 사전(

)은 훈련 신호(training signal)를 기초로 희소하게 표현된 행렬(

)로서, 테스트 신호(

)의 해를 찾을 수 있는 기능을 한다. 사전(

)은 수학식 1과 같이 나타낼 수 있다. 그리고, 수학식 1의 왼쪽 클래스에 대한 행렬(

)과 오른쪽 클래스에 대한 행렬(

)은 수학식 2와 같이 나타낼 수 있다. 여기서, 사전(

)은

차원이고,

는 훈련 신호의 개수이다.The dictionary unit 110 is a dictionary (based on the training signal)

). dictionary(

) Is a sparse representation of a matrix based on a training signal (

, The test signal (

Function to find the solution. dictionary(

) Can be expressed as in Equation 1. Then, the matrix for the left class of Equation 1

) And the matrix for the right class (

) May be expressed as in Equation 2. Where dictionary (

)silver

Dimension,

Is the number of training signals.

)을 만들 수 있다. 또는 뇌파 신호의 분류 성능을 높이기 위해, 사전부(110)는 뇌파 신호를 다양한 필터링 방법으로 필터링하여 사전(

)을 만들 수 있다. 예를 들면, 사전부(110)는 CSP(common spatial pattern) 필터링 방법, PCA(Principal component analysis) 필터링 방법, 또는 ICA(Independent component analysis) 필터링 방법 등으로 훈련 신호를 필터링할 수 있다. 여기서는 CSP 필터링 방법 위주로 설명한다. The dictionary unit 110 accumulates a training signal in a column and advances the dictionary (

). Alternatively, in order to increase the classification performance of the EEG signals, the dictionary unit 110 filters the EEG signals by various filtering methods.

). For example, the dictionary unit 110 may filter the training signal using a common spatial pattern (CSP) filtering method, a principal component analysis (PCA) filtering method, or an independent component analysis (ICA) filtering method. This section focuses on the CSP filtering method.

Here, the training signal is an electroencephalogram (EGG) signal obtained through a motion imagination experiment, and is a signal corresponding to a plurality of classes to be classified, for example, left and right classes. There are various ways to obtain training signals. For example, subjects perform BCI experiments on motor imagery of left and right hands. Subjects attached gold-plated electrodes attached to the scalp and measured for a defined number of EEG channels. When the experiment begins, if the letter indications of the left hand or the right hand appear random, the subject imagines the left or right hand movements, such as clenching fists, as directed. The collected signal is extracted from a specific time point, for example, 256 samples, after a command appears. Since the optimum start time differs for each test subject, a start time for maximizing classification accuracy can be selected for each test subject.

)을 통해서 테스트 신호(

)의 해를 푼다. 즉, 풀이부(130)는 수학식 3의 선형방정식을 풀어 해(

)를 구한다.The pool part 130 is a dictionary (

Through the test signal (

Solve the year. That is, the solver 130 solves the linear equation of equation (3) (

).

)는 테스트 신호(

)와 관계된 몇 개의 원소만 0이 아니고, 나머지 원소는 0인 희소 벡터(sparse vector)이다. 이때, 테스트 신호(

)의 희소 표현은 해(

)의 0이 아닌 원소의 개수가

보다 훨씬 적을 때 만들어진다.year(

) Is the test signal (

Are not sparse, only a few elements are associated with the zero, and the rest are zero sparse vectors. At this time, the test signal (

The rare expression of)

), The number of nonzero elements

Created when much less than.

)의 열의 크기(

)가 행(

)보다 큰 언더디터민드(underdetermined) 상태이다. 따라서, 풀이부(130)는 L1 최소화 방법(L1 norm minimization)으로 해(

)를 구하며, 수학식 4와 같이 표현된다.When the solver 130 solves Equation 3, the dictionary (

Column's size (

) Rows (

Is underdetermined greater than). Therefore, the solver 130 is a solution L1 minimization (L1 norm minimization)

) And is expressed as Equation 4.

)가 입력되는 경우, 해(

)는 수학식 5와 같이 구해진다. 이때, 테스트 신호(

)는 오른쪽 움직임 상상 신호이므로, 수학식 6과 같이 0이 아닌 계수(coefficient)가 오른쪽 클래스에 해당하는 해(

)에 더 많이 나타난다. For example, the test signal corresponding to the right motion imaginary signal (

) Is entered, the solution (

) Is obtained as in Equation 5. At this time, the test signal (

) Is a right motion imaginary signal, so a solution whose nonzero coefficient corresponds to the right class

Appears more).

)를 기초로 테스트 신호(

)를 분류한다. 이때, 최소화 문제를 풀어 획득한 해(

)는 0이 아닌 원소들을 가지는데, 이 0이 아닌 원소들은 어느 하나의 클래스에 해당해야 한다. 하지만, 뇌파 신호는 잡음이 매우 심하기 때문에, 수학식 6과 같이 0이 아닌 원소들이 다른 클래스에도 나타날 수 있다. 따라서, 분류부(150)는 특성함수(

)를 이용하여, 해(

)가 어느 클래스에 해당하는지 분류한다. 여기서, 특성함수(

)는 해(

)에서, 다른 클래스와 관련된 모든 원소들을 0으로 채운(nullify) 벡터이다. 예를 들어, 특성함수(

)는 해(

)에서,

의 원소들을 0으로 채운 벡터(

)이고, 특성함수(

)는 해(

)에서,

의 원소들을 0으로 채운 벡터(

)이다. The classification unit 150 is the solution (

Based on the test signal (

Classify In this case, the solution obtained by solving the minimization problem (

) Has nonzero elements, and these nonzero elements must belong to either class. However, since the EEG signal is very noisy, non-zero elements may appear in other classes as shown in Equation (6). Therefore, the classification unit 150 is a characteristic function (

), The solution (

) Classify which class it belongs to. Where the characteristic function (

) Is the year (

), All the elements associated with the other class It is a null filled vector. For example, the characteristic function (

) Is the year (

)in,

A vector of zero elements

) And the characteristic function (

) Is the year (

)in,

A vector of zero elements

)to be.

)를 사전(

)에 투영한 값[

]과 테스트 신호(

)의 차이를 기초로 클래스별 나머지[

]를 계산한다. 그리고, 분류부(150)는 수학식 8과 같이 나머지[

]가 최소가 되는 클래스로 테스트 신호(

)를 분류한다.The classification unit 150 is a characteristic function (Equation 7)

)

Projected to) [

] And test signal (

Rest of each class based on the difference between

]. And, the classification unit 150 is the rest [

] Is the class with the minimum test signal (

Classify).

2 is a schematic block diagram of a brain-computer connection device according to another embodiment of the present invention, FIG. 3 is a diagram showing a dictionary designed according to an embodiment of the present invention, and FIG. 4 is an embodiment of the present invention. It is a figure which shows the linear equation by.

)을 설계할 수 있다. First, referring to FIG. 2, the brain-computer connection device 100 includes a dictionary unit 110, a solver 130, a classifier 150, a CSP filtering unit 170, and feature signal extraction. A feature extraction unit 190. The dictionary unit 110 uses a training signal that has passed through the CSP filtering unit 170 and the feature signal extractor 190.

) Can be designed.

The CSP filtering unit 170 spatially filters the received training signal to maximize one class and minimize the other class. That is, the CSP filtering unit 170 makes the training signal a signal having a low correlation between two classes, that is, a left class and a right class.

를 뇌파 신호의 세그먼트로 정의한다. 이때

는 뇌파 채널 수이고, T는 한 명의 피실험자에 대한 모든 시행의 샘플 시점(time point)이다. 뇌파 훈련 신호는 왼손 움직임 상상과 오른손 움직임 상상에 각각 해당하는 두 클래스(

_,

₎를 포함한다.CSP filtering performed by the CSP filtering unit 170 is as follows. here,

Is defined as the segment of the EEG signal. At this time

Is the number of EEG channels and T is the sample time point of all trials for one subject. EEG training signals are divided into two classes, one for imagining the left hand movement and one for imagining the right hand movement.

_,

₎ .

)을 찾는다. 이때, 필터 행렬(

)의 각 열벡터(

)는

이다. CSP 필터링부(170)는 필터 행렬(

)의 앞쪽 n개의 CSP 필터와 뒤쪽 n개의 CSP 필터를 사용하여 2n CPS 필터(

)를 만든다. CSP 필터 행렬(

)은 수학식 9와 같이 나타내진다.The CSP filtering unit 170 may filter the filter matrix, which is a set of CSP filters based on the CSP algorithm.

). In this case, the filter matrix (

Each column vector ()

)

to be. The CSP filtering unit 170 is a filter matrix (

2n CPS filter (using front n CSP filters and rear n CSP filters)

). CSP filter matrix (

) Is represented by Equation (9).

,

)는 수학식 10과 같이 CSP 필터링된 신호로 정의될 수 있다.Training signals corresponding to imagining left and right hand movements (

,

) May be defined as a CSP filtered signal as shown in Equation 10.

The feature signal extractor 190 extracts a signal of a specific frequency band from the CSP filtered signal as a feature signal. The feature signal may be a signal corresponding to sensorimotor rhythms (SMRs) of the sensorimotor cortex of the scalp. The feature signal extractor 190 may extract the feature signal in various ways. The feature signal extractor 190 extracts a feature signal by calculating a band power of a specific frequency band. Alternatively, the feature signal extractor 190 may extract a coefficient of a specific frequency band obtained by FFT (fast Fourier transform) for each CSP channel as a feature signal. The specific frequency band may be part of the Mu (8-14 Hz) or Beta (15-30 Hz) band of sensorimotor rhythms (SMRs).

)을 설계한다. 도 3을 참고하면, 사전(

)은 왼쪽 클래스에 대한 행렬(

)과 오른쪽 클래스에 대한 행렬(

)로 구성된다. 그리고 사전(

)의 행의 수는 CSP 필터의 수이고, 각 클래스의 열의 수는 각 클래스의 훈련 신호의 수이다.The dictionary unit 110 uses a training signal that has passed through the CSP filtering unit 170 and the feature signal extractor 190.

). Referring to Figure 3, the dictionary (

) Is the matrix (for the left class

) And the matrix for the right class (

). And the dictionary (

The number of rows) is the number of CSP filters, and the number of columns in each class is the number of training signals in each class.

)을 통해서 테스트 신호(

)의 해를 푼다. 이때, 풀이부(130)는 L1 최소화 방법(L1 norm minimization)으로 해(

)를 구한다. 이때, 테스트 신호(

)는 사전(

)의 열을 만드는 방식과 같이 CSP 필터링부(170)와 특징 신호 추출부(190)를 거친 신호이며, 벡터화를 통해 m 차원의 벡터(

)로 변환된다. 도 4를 참고하면, 풀이부(130)는 수학식 4의 선형방정식(

)을 푼다. 만약, 오른쪽 클래스에 해당하는 테스트 신호(

)가 입력되는 경우, 해(

) 중에서 오른쪽 클래스(

)에서 0이 아닌 계수(scalar coefficient)들이 다른 클래스보다 더 나타난다.Referring back to FIG. 2, the pooling unit 130 is a dictionary (

Through the test signal (

Solve the year. At this time, the pool 130 is a solution L1 minimization (L1 norm minimization)

). At this time, the test signal (

) Is a dictionary (

Like the way to create a column of) is a signal passed through the CSP filtering unit 170 and the feature signal extraction unit 190, through the vectorization of the m-dimensional (

Is converted to). 4, the solver 130 is a linear equation (4)

) If the test signal for the right class (

) Is entered, the solution (

Of the right class (

), Non-scalar coefficients appear more than other classes.

)를 기초로 클래스별 나머지(

)를 계산하고, 나머지가 최소가 되는 클래스로 테스트 신호(

)를 분류한다.Referring back to FIG. 2, the classifying unit 150 has a characteristic function (

Based on the rest of the class (

), And the test signal (

Classify

Next, various methods for extracting the feature signal by the feature signal extractor 190 will be described.

)은 밴드파워 계산으로 추출한 각 성분(element)을 이용하여 설계된다. 즉, 사전(

)의 열은 2n개의 밴드파워로 구성될 수 있다.The feature signal extractor 190 may extract the feature signal through band power computation. In this case, the feature signal extractor 190 calculates a band moment of a specific frequency band by calculating a second moment of the CSP filtered signal. dictionary(

) Is designed using each element extracted by band power calculation. That is, the dictionary (

) May be composed of 2n band powers.

Alternatively, the feature signal extractor 190 may extract the feature signal through the FFT. In this case, the feature signal extractor 190 leaves only coefficients corresponding to a specific frequency band. The signal output from the feature signal extractor 190 may be represented by Equation (11).

는 특징 신호로 추출된 열의 개수이다. 예를 들어, 샘플링 주파수가 256 samples/sec이고, 주파수 해상도가 1Hz이며, Mu(8~14Hz)와 Beta(15~30Hz)에 해당하는 밴드파워를 추출하는 경우,

는 최대 24이다. here,

Is the number of columns extracted as the feature signal. For example, if the sampling frequency is 256 samples / sec, the frequency resolution is 1 Hz, and the band power corresponding to Mu (8-14 Hz) and Beta (15-30 Hz) is extracted,

Is up to 24.

의 2n개의 행들을 연결한 뒤 트랜스포즈(transpose)하여 벡터화하여 사전(

)의 왼쪽 행렬(

)을 설계한다. 마찬가지로, 사전부(110)는

의 2n개의 행들을 연결한 뒤 트랜스포즈하여 벡터화하여 사전(

)의 오른쪽 행렬(

)을 설계한다. In this case, the dictionary unit 110

Concatenate 2n rows of and transpose and vectorize

Left matrix ()

). Similarly, the dictionary unit 110

Concatenate 2n rows of and transpose and vectorize

Matrix on right

).

5 is a diagram illustrating a training signal relationship before CSP filtering according to an embodiment of the present invention, and FIG. 6 is a diagram showing a training signal relationship after CSP filtering according to an embodiment of the present invention.

First, referring to FIG. 5, left and right samples are correlated with a training signal before CSP filtering.

Referring to FIG. 6, training signals after CSP filtering are less relevant than before CSP filtering. In other words, since CSP filtering spatially filters the input training signal and maximizes one class and minimizes the other class, the left and right samples after CSP filtering are more related than the left and right samples after CSP filtering. Write well

)을 만들면, 불규칙(incoherent)한 사전(

)을 설계할 수 있다. 그리고, 압축센싱 시, 사전(

)이 불규칙할수록 희소 표현 기반 분류의 정확도를 높일 수 있다. 결국, 뇌-컴퓨터 접속 장치(100)는 잡음이 매우 심하고 시간에 따라 변화하여 분류가 쉽지 않은 움직임 상상 뇌파 신호를 더 정확히 분류할 수 있다.Therefore, the brain-computer connection device 100 performs CSP filtering on the training signal.

), An incoherent dictionary (

) Can be designed. And during compression sensing, the dictionary (

), The more irregular the higher the accuracy of sparse representation-based classification. As a result, the brain-computer connection device 100 may classify the motion imaginary EEG signal which is very noisy and changes over time, which is not easy to classify.

7 is a flowchart of a classification method according to an embodiment of the present invention.

Referring to FIG. 7, the brain-computer connection device 100 obtains a training signal corresponding to each class (S710). The training signal is an EEG signal obtained through a motion imagination experiment.

The brain-computer connection device 100 performs CSP filtering to maximize the training signal for one class and minimize the other class (S720). The brain-computer connection device 100 processes the CSP filtering so that the left class and the right class are reduced in the training signal so as to be distinguished.

)의 해(

)를 풀기 위한 사전(

)을 설계한다(S730). 뇌-컴퓨터 접속 장치(100)는 훈련 신호의 정보가 희소 표현된 사전(

)을 만든다. 특히, 뇌-컴퓨터 접속 장치(100)는 CSP 필터링한 신호를 기초로 불규칙(incoherent)한 사전(

)을 설계할 수 있다. 이때, 뇌-컴퓨터 접속 장치(100)는 CSP 필터링한 신호 중에서 특정 주파수 대역의 신호를 특징 신호로 추출하고, 이 특징 신호를 기초로 사전(

)을 설계할 수 있다. 즉, 뇌-컴퓨터 접속 장치(100)는 감각운동 리듬에 해당하는 신호를 특징 신호로 추출할 수 있다.Brain-computer connection device 100 is based on the CSP filtered signal test signal (

)due to(

Dictionary () to solve

) Is designed (S730). The brain-computer connection device 100 is a dictionary (in which the information of the training signal is scarcely expressed)

). In particular, the brain-computer connection device 100 is an incoherent dictionary (based on the CSP filtered signal).

) Can be designed. At this time, the brain-computer connection device 100 extracts a signal of a specific frequency band from the CSP filtered signal as a feature signal, and based on the feature signal,

) Can be designed. That is, the brain-computer connection device 100 may extract a signal corresponding to the sensory motor rhythm as a feature signal.

)를 입력받아 선형방정식(

)의 해(

)를 구한다(S740). 이때, 선형방정식은 언더디터민드(underdetermined) 상태이므로, 뇌-컴퓨터 접속 장치(100)는 L1 최소화 방법으로 해(

)를 구한다.Brain-computer connection device 100 is a test signal (

), The linear equation (

)due to(

To obtain (S740). At this time, since the linear equation is underdetermined, the brain-computer connection device 100 uses the L1 minimization method (

).

) 중에서 0이 아닌 계수의 분포를 기초로 테스트 신호(

)를 분류한다(S750). 해(

)는 테스트 신호(

)와 관계된 몇 개의 원소만 0이 아니고, 나머지 원소는 0인 희소 벡터이다. 따라서, 뇌-컴퓨터 접속 장치(100)는 왼쪽 클래스에 해당하는 해(

)의 계수들이 다른 클래스보다 더 나타나면 테스트 신호(

)를 왼쪽으로 분류하고, 오른쪽 클래스에 해당하는 해(

)의 계수들이 다른 클래스보다 더 나타나면 테스트 신호(

)를 오른쪽으로 분류한다. 이를 위해, 뇌-컴퓨터 접속 장치(100)는 해(

)에서, 다른 클래스와 관련된 모든 원소들을 0으로 채운 벡터인 특성함수(

)를 이용하여 수학식 7과 같이 클래스별 나머지를 계산한다. 그리고, 뇌-컴퓨터 접속 장치(100)는 수학식 8과 같이 나머지가 최소가 되는 클래스로 테스트 신호(

)를 분류할 수 있다.The brain-computer connection device 100 is a classed solution (

Of test signals based on the distribution of nonzero coefficients

) Are classified (S750). year(

) Is the test signal (

Only a few elements related to) are nonzero, and the rest are zero sparse vectors. Therefore, the brain-computer connection device 100 has a solution corresponding to the left class (

), If more coefficients than other classes, the test signal (

) To the left, and the solution for the right class (

), If more coefficients than other classes, the test signal (

) To the right. To this end, the brain-computer connection device 100 can solve (

), All the elements associated with the other class A characteristic function that is a vector filled with zeros (

) To calculate the remainder for each class as shown in equation (7). In addition, the brain-computer connection device 100 has a test signal in a class in which the rest is minimum, as shown in Equation (8).

) Can be classified.

FIG. 8 is a diagram illustrating an easy classification result by a linear discrimination method, and FIG. 9 is a diagram illustrating a classification result according to an embodiment of the present invention.

Referring to FIG. 8, a linear discriminant analysis (LDA) maps a right training signal (o mark) and a left training signal (x mark), respectively. The test signal is classified based on a decision line separating the right training signal and the left training signal. That is, in the linear discrimination method, when the test signal is mapped below the discrimination line, it is classified as the right imaginary signal, and when it is mapped over the discrimination line, it is classified as the left imaginary signal. At this time, if the left test signal that is easy to discriminate is input, the left test signal is mapped to a position away from the discrimination line, and the linear discrimination method easily distinguishes the test signal into the left imaginary signal.

Referring to FIG. 9, when the brain-computer connection device 100 also inputs a test signal that is easy to discriminate, non-zero coefficients appear in a nearly left solution, so that the test signal is easily classified into a left imaginary signal.

10 is a diagram illustrating a difficult classification result by a linear discrimination method, and FIG. 11 is a diagram showing a classification result according to another embodiment of the present invention.

Referring to FIG. 10, when a test signal that is difficult to discriminate is input, the linear discrimination method is mapped near the discrimination line. Thus, although the left test signal is a left signal, it may be mapped below the discrimination line. In this case, the linear discrimination method may misclassify the left test signal as the right imaginary signal. Since there is a gray region in which the linear discrimination method is difficult to discriminate, there is a limit in classifying the test signal accurately.

Referring to FIG. 11, when a test signal that is difficult to discriminate is input to the brain-computer connection device 100, coefficients other than zero appear in the right solution as well as the left solution. However, since the brain-computer connection device 100 calculates the sparse vector based on the sparse representation of the test signal, the coefficients shown on the left are more represented than the coefficients shown on the right. Therefore, the brain-computer connection device 100 is more likely to classify correctly even if a difficult test signal is input.

The embodiments of the present invention described above are not implemented only by the apparatus and method, but may be implemented through a program for realizing the function corresponding to the configuration of the embodiment of the present invention or a recording medium on which the program is recorded.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is to be understood that the invention is not limited to the disclosed exemplary embodiments, It belongs to the scope of right.

Claims (20)

A brain-computer connection device that classifies a test signal, an EEG signal, into one of a left class and a right class,
The training signal of the left class and the training signal of the right class are received, and the training signal of each class is filtered to extract a first matrix corresponding to the left class and a second matrix corresponding to the right class. A dictionary unit for designing a dictionary by accumulating a matrix and the second matrix,
A solution unit for solving a test signal based on the dictionary, and
A classification unit classifying the test signal into one of the left class and the right class based on a distribution of nonzero coefficients represented in each class of the solution.
Lt; / RTI >
Wherein each of the first matrix and the second matrix is a matrix sparsely represented by the filtering. In claim 1,
The apparatus further includes a CSP filtering unit configured to output a CSP filtered signal by performing CSP (common spatial pattern) filtering to maximize the training signal for one class and minimize the other class.
And the dictionary unit designs the dictionary based on the CSP filtered signal. 3. The method of claim 2,
The dictionary part
The number of rows is the number of CSP filters, and the number of columns of each class is the number of training signals of each class. 3. The method of claim 2,
Further comprising a feature signal extraction unit for extracting a signal of a specific frequency band from the CSP filtered signal as a feature signal,
And the dictionary unit designs the dictionary based on the feature signal. 5. The method of claim 4,
Wherein said specific frequency band is a frequency band corresponding to the sensory rhythm of the sensorimotor cortex portion of the scalp. In claim 1,
The glue part
A brain-computer connection device that solves the solution by L1 minimization. In claim 1,
The classifier
A brain-computer connection device for classifying the test signal into a class in which the most non-zero coefficients appear among each class of the solution. In claim 1,
The classifier
Calculate the remainder for each class based on the solution and classify the test signal into a class having the least remainder,
The remainder for each class is a value calculated by the difference between a first value projecting a vector generated by filling an element of a class other than itself with zero in the solution and a second value corresponding to the test signal. Device. In claim 1,
And the training signal and the test signal are motion imaginary based brain wave signals. In claim 1,
The left class is a class corresponding to the left movement imagination, and the right class is a class corresponding to the right movement imagination. The brain-computer connection device classifies a test signal, which is an EEG signal, into one of a left class and a right class,
Acquiring a training signal corresponding to the left class and a training signal corresponding to the right class;
Filtering the training signals of each class to maximize the corresponding class and minimize the other class, extracting a first matrix corresponding to the left class and a second matrix corresponding to the right class;
Designing a dictionary by accumulating the first matrix and the second matrix;
Solving the test signal based on the dictionary, and
Classifying the test signal into one of the left class and the right class based on a distribution of nonzero coefficients represented in each class of the solution.
Lt; / RTI >
And the first matrix and the second matrix are matrices sparsely represented by the filtering. 12. The method of claim 11,
And the training signal and the test signal are motion imaginary based brain wave signals. 12. The method of claim 11,
And the left class is a class corresponding to a left movement imagination, and the right class is a class corresponding to a right movement imagination. 12. The method of claim 11,
The filtering is Common Spatial Pattern (CSP) filtering,
The step of designing the dictionary
And the number of rows is the number of CSP filters and the number of columns of each class is the number of training signals of each class. The method of claim 14,
Extracting a signal of a specific frequency band from the CSP filtered signal as a feature signal;
Designing the dictionary comprises designing the dictionary based on the feature signal. 16. The method of claim 15,
The specific frequency band is a frequency band corresponding to the sensory motor rhythm of the sensorimotor cortex portion of the scalp. 12. The method of claim 11,
The solution step is
A classification method for solving the solution by L1 minimization. 12. The method of claim 11,
The solution step is
A classification method for obtaining the solution that is so small that the number of nonzero coefficients can be determined to be less than the number of training signals of each class. 12. The method of claim 11,
The classifying step
A classification method for classifying the test signal into classes in which the most non-zero coefficients appear among the classes of the solution. 12. The method of claim 11,
The classifying step
The test signal is classified into a class having a least residual by calculating a residual for each class based on the solution, and the residual for each class projects a vector generated by filling an element of a class other than itself with zero in the solution. And a value calculated by a difference between one first value and a second value corresponding to the test signal.

Images