CN103258215A - Multi-lead correlation analysis electroencephalo-graph (EEG) feature extraction method - Google Patents

Abstract

The invention relates to an EEG feature extraction method for correlation analysis among multiple leads. In multi-type motor imagery task recognition, effectively extracting the EEG signal features of brain regions activated by specific motor imagery tasks is one of the key issues to improve the recognition rate. The present invention first collects multi-lead motor imagery EEG signals, and secondly obtains the correlation coefficient matrix for the correlation coefficients between each lead EEG signal, and then calculates the difference between the row variance of the correlation coefficient matrix and the variance sum of all rows. The ratio and its natural logarithm, the obtained result is used as the feature vector of the EEG signal, and finally the feature vector is input into the classifier to complete the classification and recognition of multi-category motor imagery tasks. The method proposed by the present invention can not only completely extract the EEG signal features on multiple brain regions simultaneously activated by specific motor imagery tasks, but also greatly reduce the influence of individual differences in EEG signals on the characteristic parameters, and can overcome the problem of electrode Under-selected questions.

Description

An EEG Feature Extraction Method for Correlation Analysis Between Multiple Leads

technical field

The invention belongs to the field of electroencephalogram signal processing, and relates to an electroencephalogram signal feature extraction method, in particular to a feature extraction method for multi-lead motor imagery electroencephalogram signals.

Background technique

Electroencephalogram (electroencephalogram, EEG) is the potential change caused by the synaptic transmission signal of the cerebral cortex nerve cell group, which can reflect the brain's autonomous or induced consciousness activities, and is closely related to the actual action behavior. Since German scientist Hans Berger recorded the electrical activity of the human brain in 1929, people have been trying to interpret human thinking activities through the identification of EEG signals. The brain-computer interface (BCI) supported by it is considered to be an important milestone in the process of human understanding of the brain. BCI does not rely on the participation of muscles and peripheral nerves, and directly realizes the communication between the human brain and the computer, which is one of the current international frontier research hotspots.

节律的阻断现象，选择脑电信号中的

节律作为开关控制信号，通过睁闭眼触发轮椅的方向开关实现与轮椅的人机交互，实验验证了系统的可行性，同时指出主观因素对

节律的阻断影响以及环境噪声对控制性能的影响还有待进一步研究，使用时需要眼部肌肉协同作用。Pfurtscheller领导的研究中心研究表明，单边的肢体运动或者仅是想象运动能激活主要的感觉运动皮层，大脑对侧产生

和

节律的ERD，同侧产生

和

节律的ERS。随后结合虚拟现实，一个C4/C5级脊髓损伤患者能够用想象运动控制轮椅在虚拟街道上运动，实验表明平均成功率为90%。日本Tanaka等设计了一种基于运动想象脑电信号控制电动轮椅的实验系统，控制轮椅从目标A正确地移动到了目标B。瑞士、西班牙和比利时的研究团队针对脑电信号的非稳定性特点以及当前自适应解决方法的缺点，研究了异步非侵入式BCI并用于轮椅控制，在稳定的脑电特征选择和共享控制上取得了较好的研究成果，实验结果表明在虚拟、实际环境下受试者都能通过想象左手运动、休息和单词关联三类任务通过异步非侵入式BCI接口分别控制轮椅左转、前进及右转运动，虚拟环境下目标到达率为100%，实际环境下目标到达率为80%，同时指出训练负担较重，分类精度有待提高。清华大学高上凯研究组开展了基于运动想象的光标移动、康复辅助训练、机器狗踢足球等BCI系统研究，在想象运动系统实验中，10名受试者通过想象左右手、脚运动产生脑电信号对系统进行控制，实验结果表明想象左右手两类分类任务的在线和离线分析平均正确率为94.92%和92.86%，想象左右手和脚三类分类任务的在线和离线分析平均正确率为85%和79.48%，同时指出差别较小的思维任务很难从空间分辨率较低的脑电信号来提取，增加系统可识别任务的种类通常会直接导致识别正确率的下降。 So far, people still know little about the formation process of brain thinking, and it is not realistic to read people's various thinking activities through EEG. However, some progress has been made in the use of EEG signals for motor imagery recognition. Pfurtscheller from the University of Graz in Austria and Wolpaw from the Wadsworth Research Center in the United States have done a lot of work on motion pattern recognition through motor imagery. Research has proved that motor imagery and actual motion will cause the same neuron activity in the main sensorimotor functional area of the brain, through EEG signals recorded by electrodes placed in sensorimotor areas allow for motor imagery pattern recognition. Wan Baikun's research group at Tianjin University uses vision to centralize control

Rhythm blocking phenomenon, select the EEG signal

The rhythm is used as a switch control signal, and the human-computer interaction with the wheelchair is realized by triggering the direction switch of the wheelchair by opening and closing the eyes. The experiment verifies the feasibility of the system, and points out that the subjective factors

The effect of rhythm blockade and the effect of environmental noise on control performance needs to be further studied, and the synergy of eye muscles is required for use. Studies at the research center led by Pfurtscheller have shown that unilateral limb movement or just imagined movement activates the main sensorimotor cortex, and the opposite side of the brain produces

and

Rhythmic ERD, arising ipsilaterally

and

Rhythmic ERS. Then combined with virtual reality, a patient with C4/C5 spinal cord injury can use imaginative motion to control a wheelchair to move on a virtual street. Experiments show that the average success rate is 90%. Japan Tanaka et al. designed an experimental system for controlling an electric wheelchair based on motor imagery EEG signals, and controlled the wheelchair to move from target A to target B correctly. Research teams from Switzerland, Spain and Belgium aimed at the instability characteristics of EEG signals and the shortcomings of current adaptive solutions, studied asynchronous non-invasive BCI and used it for wheelchair control, and made achievements in stable EEG feature selection and shared control. The experimental results show that in the virtual and real environments, the subjects can control the wheelchair to turn left, forward and right through the asynchronous non-invasive BCI interface by imagining the three tasks of left hand movement, rest and word association. Sports, the target arrival rate in the virtual environment is 100%, and the target arrival rate in the actual environment is 80%. At the same time, it is pointed out that the training burden is heavy and the classification accuracy needs to be improved. Gao Shangkai’s research group at Tsinghua University has carried out research on BCI systems based on motor imagination, such as cursor movement, rehabilitation assistance training, and robot dogs playing football. The signal controls the system. The experimental results show that the average accuracy rates of online and offline analysis of the classification tasks of imagining left and right hands are 94.92% and 92.86%, and the average accuracy rates of online and offline analysis of classification tasks of imagining left and right hands and feet are 85% and 85%. 79.48%. At the same time, it is pointed out that thinking tasks with small differences are difficult to extract from EEG signals with low spatial resolution. Increasing the types of tasks that can be recognized by the system usually directly leads to a decline in the recognition accuracy.

A comprehensive study at home and abroad found that the single spontaneous EEG generated by motor imagery does not require external stimulation signals, and is the source of body movement. The brain bioelectric information collected through EEG signals contains the control information of brain motor imagery. However, there are still some major problems in the study of motor imagery EEG: first, the resolution of EEG signals extracted through thinking tasks is low, especially for motor imagery tasks with small differences; second, increasing the types of recognition tasks will directly lead to recognition decline in accuracy. One of the key issues affecting the recognition rate is to effectively extract the features corresponding to different motor imagery tasks from weak background noise, random and non-stationary EEG signals. Researchers use various methods to extract effective EEG features, such as Fourier transform, autoregressive model, power spectrum and adaptive regression model, fourth-order cumulant, wavelet transform, wavelet packet transform, Hilbert- Huang transform, complexity analysis method, tensor analysis method, public space mode, etc., and then identify different motor imagery tasks, and achieved rich research results. However, the current EEG feature extraction methods only analyze the information of a small number of channels. The benefits of doing so are obvious, requiring fewer electrodes, which not only shortens the preparation time, but also requires a small amount of information processing costs for a small amount of data. Correspondingly, scholars such as Blankertz, Sannelli, Schroder, and Barachant pointed out that a small number of channels selected with neurophysiological prior knowledge does not necessarily produce better results than full-channel acquisition, and insufficient selection of electrodes will also reduce the classification accuracy. How to completely extract the EEG signal features of multiple brain regions simultaneously activated by specific motor imagery tasks for different subjects is exactly the problem that the present invention tries to solve.

Although different areas of the cerebral cortex perform relatively independent functions, the completion of a specific motor imagery task requires the simultaneous participation of one or several spatially separated functional areas, and different motor imagery tasks are performed at the same time. The activated motor cortex regions are also different.

Contents of the invention

The purpose of the present invention is to provide a method for extracting EEG features based on correlation analysis between multiple leads to address the shortcomings of existing EEG feature extraction methods.

For multi-lead EEG signals, the area measured by the electrodes corresponding to each lead (channel) is often defined as a node, and its electrical activities are several time series. First calculate the correlation coefficient between these time series to obtain the correlation coefficient matrix, then calculate the ratio between the row variance of the correlation coefficient matrix and the variance sum of all rows and its natural logarithm as the eigenvector of the EEG signal, and finally Multiple classes of motor imagery tasks were identified using these features.

In order to achieve the above object, the inventive method mainly comprises the following steps:

Step (1) Obtain multi-channel motor imagery EEG signal sample data. First, the motor imagery EEG signal was collected by multi-lead electrode cap, and then preprocessed by band-pass filtering method.

相关系数矩阵

，  Step (2) Correlation coefficient calculation. According to the time series similarity measurement method shown in formula (1), the correlation coefficient between each lead (channel) EEG signal is calculated, and a

correlation coefficient matrix

,

          (1)

(1)

和

为通道

和

在时刻的脑电信号数值，

代表通道

与通道

之间的相关系数值，

为时间序列长度，

表示采集脑电信号的通道数。 in,

and

for the channel

and

exist The value of the EEG signal at the moment,

representative channel

with channel

The correlation coefficient value between,

is the length of the time series,

Indicates the number of channels for collecting EEG signals.

Step (3) EEG feature extraction. On the basis of the correlation coefficient matrix in step (2), calculate the variance of each row of the matrix and the sum of variances of all rows, and then calculate the eigenvector of the EEG signal according to formula (2)

                (2)

(2)

)表示相关系数矩阵的每一行，

表示自然对数运算，

表示方差运算。 in, (

) represents each row of the correlation coefficient matrix,

represents the natural logarithm operation,

Represents the variance operation.

Compared with the existing motor imagery EEG feature extraction algorithm, the method proposed in the present invention can not only completely extract the EEG signal features of multiple brain regions simultaneously activated by a specific motor imagery task, but also greatly reduce the EEG signal characteristics. The influence of individual differences on feature extraction parameters can overcome the problem of insufficient electrode selection, and it is simple and easy to implement.

The method of the invention can better meet the feature extraction requirements in multi-pattern recognition tasks, and has broad application prospects in the fields of brain-computer interface and brain disease diagnosis.

Description of drawings

Fig. 1 is the implementation flowchart of the present invention.

Detailed ways

The EEG characteristic method based on the correlation analysis among multiple leads of the present invention will be described in detail below in conjunction with the accompanying drawings, and FIG. 1 is an implementation flow chart.

As shown in Figure 1, the implementation of the method of the present invention mainly includes the following steps: (1) Acquiring multi-channel motor imagery EEG signal sample data, including the collection and preprocessing of EEG signals under several motor imagery experimental paradigms; (2) Calculate the correlation coefficient between each pair of EEG signals according to the time series similarity measurement method; (3) Calculate the ratio between the row variance of the correlation coefficient matrix and the variance sum of all rows and its natural logarithm, and convert the obtained The results are used as discriminative features to describe EEG signals; (4) Input EEG features into support vector machine classifiers for training and testing, and complete the classification of various motor imagery tasks.

Each step will be described in detail below one by one.

Step 1: Obtain multi-channel motor imagery EEG signal sample data

The 40-conductive electrode cap in the Scan4.3 acquisition device of American Neuro Scan Company was used to collect the EEG signals during the motor imagery process. The subjects sat in the wheelchair after wearing the EEG cap as required, kept quiet and natural, and watched the scene prompts set in the experimental environment. The following motor imagery experiment paradigms were used: the right hand controls the wheelchair control lever forward, the left hand controls the wheelchair control lever backward, the left foot hops and both hands push the wheelchair to move to the left, the right foot hops and both hands push the wheelchair to move to the left , respectively corresponding to the control motion forms of wheelchair forward, brake, left turn, and right turn. During the implementation process, the design of the experimental mode can be appropriately modified according to the specific conditions of the experiment. After collecting the data, the band-pass filtering method is used for signal preprocessing.

Step 2: Calculation of correlation coefficient

相关系数矩阵

，  According to time series similarity measurement methods such as Euclidean distance and Mahalanobis distance, the correlation coefficient between each pair of EEG signals was calculated. The present invention uses the Euclidean distance to measure the similarity between the time series of the EEG signals of two leads. According to the formula (1), calculate the correlation coefficient between each pair of EEG signals, and get a

correlation coefficient matrix

,

          (1)

(1)

和

为通道

和

在时刻的脑电信号数值，代表通道

与通道

之间的相关系数值，

为时间序列长度，表示采集脑电信号的通道数。 in,

and

for the channel

and

exist The value of the EEG signal at the moment, representative channel

with channel

The correlation coefficient value between,

is the length of the time series, Indicates the number of channels for collecting EEG signals.

Step 3: EEG feature extraction

的基础上，分别计算中每一行的方差以及所有行的方差和，然后根据公式(2)计算计算脑电信号的特征向量 In step two the correlation coefficient matrix

On the basis of The variance of each row and the sum of variances of all rows, and then calculate the eigenvector of the EEG signal according to formula (2)

                (2)

(2)

(

)表示相关系数矩阵的每一行，

表示自然对数运算，

表示方差运算。 in,

(

) represents each row of the correlation coefficient matrix,

represents the natural logarithm operation,

Represents the variance operation.

Step 4: Motor imagery task classification based on support vector machine

The EEG feature vector obtained in step 3 is used as the input of the support vector machine classifier for training and testing to complete the classification of various motor imagery tasks.

Claims (1)

1. the brain electrical feature extracting method of correlation analysis between a multi-lead is characterized in that this method comprises the steps:

Step (1) is obtained hyperchannel motion imagination EEG signals sample data, specifically: at first adopt multi-lead electrode cap collection campaign imagination EEG signals, adopt band-pass filtering method to carry out pre-service then;

Step (2) related coefficient is calculated, specifically: calculate the EEG signals related coefficient between any two of respectively leading according to the time series similarity measure method shown in the formula (1), obtain one Correlation matrix

,

(1)

Wherein,

With

Be passage

With

EEG signals numerical value constantly,

Represent passage

With passage

Between facies relationship numerical value,

Be length of time series,

The port number of EEG signals is gathered in expression;

Step (3) brain electrical feature extracts, specifically: on the basis of step (2) correlation matrix, respectively the variance of variance of each row of compute matrix and all row and, calculate the proper vector of EEG signals then according to formula (2)

(2)

Wherein,

Each row of expression correlation matrix,

The computing of expression natural logarithm,

The computing of expression variance.

Images