CN112733727A - Electroencephalogram consciousness dynamic classification method based on linear analysis and feature decision fusion - Google Patents

Abstract

The invention provides a linear analysis-based electroencephalogram consciousness dynamic classification method based on feature decision fusion, wherein an electroencephalogram signal data set X (X1, X2, … and Xn) is acquired through a brain wave induction helmet, and n is a positive integer; and classifying the signal data set X ═ X (X1, X2, …, Xn) by using a canonical discriminant analysis RDA and a quadratic discriminant analysis QDA to obtain a correlation coefficient matrix rho_RDAAnd ρ_QDAMeanwhile, a feature decision fusion comprising a feature extraction unit, a projection classification unit and a decision selection unit is constructed to carry out feature integration and decision selection on the decisions and coefficients of the RDA and the QDA, so that better classification accuracy is obtained. The invention integrates two algorithms by constructing feature decision fusion, thereby selecting the decision which is more likely to be accurate, and obtaining better classification accuracy rate on the classification of motor imagery data.

Description

A dynamic classification method of EEG consciousness based on feature decision fusion based on linear analysis

technical field

The invention belongs to the field of EEG dynamic analysis, in particular to a dynamic classification method of EEG consciousness based on linear analysis and feature decision fusion.

Background technique

MI is a kind of EEG signal, which reflects the activation of specific functional areas of the brain when a person is in motor imagery, and the corresponding EEG signal will produce stable and regular characteristic changes. Base. In order to decode the intention of the subject from the MI signal, various methods have been proposed to identify and classify the MI signal, such as linear discriminant analysis (LDA), Gaussian classifier, probabilistic neural network (probabilistic NN), etc. ). Especially in the method based on LDA, when classifying new samples, they are projected onto the same straight line, and then the class of the new samples is determined according to the location of the projected points. The analytical solution based on the generalized eigenvalue problem can be directly obtained, thereby avoiding the local minimum problem often encountered in the construction of general nonlinear algorithms. It is not necessary to artificially encode the output category of the pattern, so that the LDA pair is unbalanced. The handling of pattern classes presents a particularly clear advantage. Compared with neural network methods, LDA does not need to adjust parameters, so there are no problems such as learning parameters and optimization weights and selection of neuron activation functions; it is not sensitive to normalization or randomization of patterns, which is based on gradient descent. It is more prominent among the various algorithms. LDA is a classic linear learning method, which was first proposed by Fisher in 1936 on the binary classification problem, also known as Fisher's linear discrimination. The idea of linear discrimination is very simple: given a set of training samples, try to project the samples onto a straight line, so that the projection points of similar samples are as close as possible, and the projection points of different samples are as far away as possible; When classifying, project it onto the same line, and then determine the class of the new sample according to the position of the projected point. QDA is a variant of LDA in which a single covariance matrix is estimated for each class of observations. QDA is especially useful if it is known in advance that individual classes exhibit different covariances. The disadvantage of QDA is that it cannot be used as a dimensionality reduction technique. Since QDA estimates the covariance matrix for each class, it has more effective parameters than LDA. RDA is a compromise between LDA and QDA, and since RDA is a regularization technique, it is more applicable when there are many potentially relevant features. Since these methods provide different decisions, fusing multiple methods to integrate different decisions is a feasible approach to improve the overall classification accuracy.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a dynamic classification method of EEG consciousness based on feature decision fusion based on linear analysis, by integrating two algorithms with similar precision in the classification of motor imagery data, and fuse their decisions through the algorithm, thereby selecting a more accurate classification method. Accurate decision-making, so as to improve the classification accuracy of EEG signals.

A kind of EEG dynamic classification method based on linear analysis feature decision fusion of the present invention, it comprises the following steps:

S1. Collect the EEG signal data set X=(X1, X2,..., Xn) through the brain wave sensing helmet;

S2, use regular discriminant analysis RDA to classify the signal data set X=(X1, X2, . . . , Xn) to obtain a correlation coefficient matrix ρ _n , which is defined as ρ _RDA ;

是权重的矩阵转置,X是数据值，

是数据值平均数；in,

is the matrix transpose of the weights, X is the data value,

is the mean of the data values;

S3, use quadratic discriminant analysis QDA to classify signal data set X=(X1, X2,..., Xn), obtain correlation coefficient matrix ρ _n , be defined as ρ _QDA ;

是权重的矩阵转置,X是数据值，

是数据值平均数；in,

is the matrix transpose of the weights, X is the data value,

is the mean of the data values;

S4. Build a feature decision fusion device to perform feature integration and decision selection on the decisions and coefficients of RDA and QDA, and obtain dynamic classification of EEG consciousness. The specific steps are as follows;

S41. Construct a feature decision fuser. The feature decision fuser includes a feature extraction unit, a projection classification unit and a decision selection unit:

S42, feature extraction is performed on the correlation coefficient of RDA and QDA by the feature extraction unit, and a feature vector F is generated;

的比值判断，值越大说明ρ就越大，根据表达式：according to

The ratio judgment of , the larger the value, the larger the ρ, according to the expression:

Among them, w _X is the projection in the x direction, w _X ^T is the transposed projection, w _Y is the projection in the y direction, w _Y ^T is the transposed projection, the X matrix is the abscissa, X ^T is the transposed matrix, Y Matrix as abscissa, Y ^T is the transposed matrix, E[w _X ^T XY ^T w _Y ] represents the expectation of w _X ^T XY ^T w _Y , E[w _X ^T XX ^T w _X ] represents w _X ^T XX ^T w The expectation of _X , E[w _Y ^T XY ^T w _Y ] represents the expectation of w _Y ^T XY ^T w _Y ;

Obtain the largest correlation coefficient and the second largest correlation coefficient of RDA and QDA, respectively;

According to the obtained maximum correlation coefficient and the second largest correlation coefficient of the RDA and QDA algorithms, the feature vector F is generated:

是QDA的最大相关系数，

是QDA的第二大相关系数，

是RDA的最大相关系数，

是RDA的第二大相关系数；in,

is the maximum correlation coefficient of QDA,

is the second largest correlation coefficient of QDA,

is the maximum correlation coefficient of RDA,

is the second largest correlation coefficient of RDA;

S43. Divide the feature vector F into two categories: RDA-false and QDA-false through the projection classification unit;

The projection classification unit uses the linear SVM classifier to project within the range of the soft boundary objective function, and projects the feature vector F into a scalar value.

是线性核函数；Among them, v _j , j=1,2,...,N is the support vector, used to determine the maximum edge plane of the classifier, a _j >, is a parameter that can be changed and adjusted, y _j refers to the category of the j-th support vector , F is the feature vector, b is the bias,

is a linear kernel function;

The soft boundary objective function is:

Among them, _δj is the slack variable, indicating whether the sample _vj is within the edge, the degree of adjustment required, C is the adjustment coefficient that controls the trade-off between width and misclassification, and the slack variable is used to determine whether the point is within the range.

By solving the soft boundary objective function, the maximum value 2/||W|| is obtained, which is used as the boundary line. According to the position of the scalar value projection point, the features in the feature vector F are divided into RDA-false and QDA-false. kind;

S44, select the decision output of the RDA or QDA algorithm according to the classification result by the decision selection unit;

The decision selection unit is:

If an RDA-false result is obtained, the module outputs the QDA decision; otherwise, the module outputs the RDA decision to obtain a dynamic classification of EEG consciousness with better accuracy.

Preferably, in step S1, a brain wave sensing helmet is used to collect EEG signals for Emotiv;

Preferably, the solving step of ρ _RDA in step S2 is as follows;

Assign the dataset X=(X1,X2,...,Xn) to one of the K groups of classes. In the training data, the class of the data is known, so the prior probability and mean of class k are:

是类别k的数值的平均数；where ω is the total number of samples, ω _k is the number of samples in category k, X _n is the sample point,

is the mean of the values of category k;

Regular discriminant analysis RDA improves the effects of multicollinearity by modifying singular covariance values; the sample covariance for each class is estimated as follows:

是类别k的平均值；where ω is the total number of samples, X _n is the sample point,

is the mean of category k;

The covariance matrix is further adjusted by introducing a shrinkage parameter γ:

where λ is the regularization parameter, 0≤λ≤1, p is the dimension of the independent variable, I is the identity matrix, and γ is the shrinkage parameter.

The optimization objective is J(W):

这就是QDA欲最大化的目标，J的最大值为矩阵

的最大特征值，而对应的为

的最大特征值对应的特征向量。求解得到

即为确定的最佳投影方向，W^T是转置的投影，将训练集中的样本向w方向投影得到：The above formula is the generalized Rayleigh quotient of _Sk and S, where _Sk = ω _k ∑ _k ,

This is the goal that QDA wants to maximize, and the maximum value of J is the matrix

The largest eigenvalue of , and the corresponding

The eigenvector corresponding to the largest eigenvalue of . Solve to get

That is, the determined optimal projection direction, W ^T is the transposed projection, and the samples in the training set are projected to the w direction to obtain:

y=w ^T X (6)

是权重的矩阵转置,X是数据值，

是数据值平均数；in,

is the matrix transpose of the weights, X is the data value,

is the mean of the data values;

Preferably, the solving step of ρ _QDA in step S3 is as follows;

Assuming that the sample data set X = (X1, X2,..., Xn) obeys the multivariate Gaussian distribution, μ _i is the mean vector, expressed as:

Compute the sample covariance matrix ∑j:

Among them, j=1,2;

The intra-class divergence matrix S _w can be obtained as:

At the same time, the inter-class divergence matrix S _b is defined as:

S _b =(μ ₁ -μ ₂ )(μ ₁ -μ ₂ ) ^T (12)

The optimization objective is J(W):

The above formula is the generalized Rayleigh quotient of S _w and S _b , which is the goal of QDA to maximize. The maximum value of J is the largest eigenvalue of the matrix S _w ^-1 S _b , and the corresponding is S _w ^-1 S _b The eigenvector corresponding to the largest eigenvalue of . Solve to get w=S _w ^-1 (μ ₁ -μ ₂ ), which is the determined optimal projection direction, and project the samples in the training set to the w direction to get:

y=w ^T X (14)

The effect of the present invention is as follows:

1. The decision-making protocol of two different methods is integrated, which reduces the problems of low accuracy and poor adaptability of a single decision;

2. Integrating their decisions is an effective way to improve the overall performance. Based on this idea, two LDA-based algorithms are integrated to improve the classification accuracy of EEG signals.

Description of drawings

Fig. 1 is the schematic diagram of the dynamic classification method of EEG consciousness based on the feature decision fusion of linear analysis of the present invention;

Fig. 2 is the test and training schematic diagram of decision fusion of the present invention;

FIG. 3 is an overall technical roadmap of the present invention.

Detailed ways

Hereinafter, embodiments of the present invention will be described with reference to FIGS. 1 to 3 .

A kind of EEG dynamic classification method based on linear analysis and feature decision fusion of the present invention, the overall flow chart is shown in Figure 3, and its steps are as follows:

S1. Collect the EEG signal data set X=(X1, X2, . . . , Xn) through the brain wave sensing helmet, and n is a positive integer;

A dynamic task model was implemented in a virtual environment where subjects indirectly controlled the ball by applying force to the bowl, and the ball could escape. The test was carried out in a room with good sound insulation, and the experimental equipment used an Emotiv helmet to collect 14 subjects (AF3, F7, F3, FC5, T7, P7, O1, O2, P8, T8, FC6, F4, F8, AF4) The electroencephalogram signal, the electrode distribution adopts 10-20 international standard lead positioning, and the sampling frequency is 128Hz. The test data is transmitted to the computer through the USB interface. A total of 10 (6 male, 4 female) healthy participants participated in the trial, exclusion criteria were any history of visual, neurological or psychiatric disorders or any existing medication, and all subjects read and signed informed consent.

During the experiment, the subjects were required to complete the boundary avoidance task, which is a highly difficult sensorimotor task. The subjects were required to manipulate a virtual bowl with a small ball within a prescribed range. The task is successful if the bounding box is within the bounds and the ball does not overflow the bowl during the whole process; if the bowl exceeds the bounding bounds from the left side, or the ball overflows from the bowl during this process, the task fails. Bowl - The initial position of the ball is on the left. According to the characteristics of this experiment, the following two groups of data were intercepted for analysis and processing. Suppose a leftward force is applied to the bowl for a period of time. Before the force is applied, the bowl moves to the right, and the ball moves to the left relative to the bowl; due to inertia, the bowl will move to the left after a period of time. , respectively intercept the data of the first 1s and the last 1s in this period of time, and vice versa. The whole experiment includes 120 test value trials, each of which is 60 for the left and right hands, and each subject obtains a set of observed measurement values X=(X1, X2, . . . , Xn), where n=120.

S2, use regular discriminant analysis RDA to classify the signal data set X=(X1, X2, . . . , Xn) to obtain a correlation coefficient matrix ρ _n , which is defined as ρ _RDA ;

Both LDA and QDA are boundary discriminative methods that aim to find boundaries that separate sample groups or categories. The boundary divides the space into regions, each region is divided according to different groups or classes, and it also depends on the type of classifier: i.e. LDA obtains a linear boundary, where lines or hyperplanes divide the variable space into regions; QDA obtains two Secondary boundaries, where quadric surfaces divide variable space into regions. LDA assumes that all classes have a single covariance matrix, and QDA assumes that each class has a different covariance matrix. QDA allows to distinguish classes with significantly different class-specific covariance matrices and form a separate variance model for each class, while clusters represent multivariate normal distributions with the same mean. RDA is a compromise between LDA and QDA, and since RDA is a regularization technique, it is more applicable when there are many potentially relevant features. Assign the dataset X=(X1,X2,...,Xn) to one of the K groups of classes. In the training data, the class of the data is known, so the prior probability and mean of class k are:

是类别k的数值的平均数。where ω is the total number of samples, ω _k is the number of samples in category k, X _n is the sample point,

is the mean of the values of category k.

Regular discriminant analysis RDA improves the effects of multicollinearity by modifying singular covariance values. The sample covariance for each class is estimated as follows:

是类别k的平均值。where ω is the total number of samples, X _n is the sample point,

is the mean of class k.

The covariance matrix is further adjusted by introducing a shrinkage parameter γ:

where λ is the regularization parameter, 0≤λ≤1, p is the dimension of the independent variable, I is the identity matrix, and γ is the shrinkage parameter.

The optimization objective is J(w)

这就是QDA欲最大化的目标，J的最大值为矩阵S_k ^-1S的最大特征值，而对应的为S_k ^-1S的最大特征值对应的特征向量。求解得到w＝S_k ^-1(μ₁，-μ₂)，即为确定的最佳投影方向，W^T是转置的投影，将训练集中的样本向w方向投影得到：The above formula is the generalized Rayleigh quotient of _Sk and S, where _Sk = ω _k ∑ _k ,

This is the goal that QDA wants to maximize. The maximum value of J is the maximum eigenvalue of the matrix S _k ^-1 S, and the corresponding eigenvector is the maximum eigenvalue of the S _k ^-1 S. Solve to get w=S _k ^-1 (μ ₁ , -μ ₂ ), which is the determined optimal projection direction, W ^T is the transposed projection, and the samples in the training set are projected to the w direction to obtain:

y=w ^T X (6)

是权重的矩阵转置,X是数据值，

是数据值平均数；in,

is the matrix transpose of the weights, X is the data value,

is the mean of the data values;

S3, use quadratic discriminant analysis QDA to classify signal data set X=(X1, X2,..., Xn), obtain correlation coefficient matrix ρ _n , be defined as ρ _RDA ;

Quadratic Discriminant Analysis (QDA) aims to find the transformation of the input features that best distinguishes the classes in the dataset. Use quadratic discriminant analysis QDA to classify the signal data set X=(X1, X2, . . . , Xn) to obtain a correlation coefficient matrix ρ _n , which is defined as ρ _QDA ;

Assuming that the sample data set X = (X1, X2,..., Xn) obeys the multivariate Gaussian distribution, μ _i is the mean vector, as follows:

Calculate the covariance matrix ∑j of the sample as:

Among them, j=1,2;

The intra-class divergence matrix S _w can be obtained as:

At the same time, the inter-class divergence matrix S _b is defined as:

S _b =(μ ₁ -μ ₂ )(μ ₁ -μ ₂ ) ^T (12)

The optimization objective is J(w)

The above formula is the generalized Rayleigh quotient of S _w and S _b , which is the goal of QDA to maximize. The maximum value of J is the largest eigenvalue of the matrix S _w ^-1 S _b , and the corresponding is S _w ^-1 S _b The eigenvector corresponding to the largest eigenvalue of . Solve to get w=S _w ^-1 (μ ₁ -μ ₂ ), which is the determined optimal projection direction, and project the samples in the training set to the w direction to get

y=w ^T X (14)

是权重的矩阵转置,X是数据值，

是数据值平均数；in,

is the matrix transpose of the weights, X is the data value,

is the mean of the data values;

S4. Build a feature decision fusion device to perform feature integration and decision selection on the decisions and coefficients of RDA and QDA, and obtain dynamic classification of EEG consciousness. The specific steps are as follows;

S41. Construct a feature decision fuser. The feature decision fuser includes a feature extraction unit, a projection classification unit and a decision selection unit:

S42, feature extraction is performed on the correlation coefficient of RDA and QDA by the feature extraction unit, and a feature vector F is generated;

的比值判断，值越大说明ρ就越大，根据如下表达式分别得到RDA和QDA的最大相关系数和第二大相关系数；according to

The larger the value, the larger the ρ, and the maximum correlation coefficient and the second largest correlation coefficient of RDA and QDA are obtained according to the following expressions;

Among them, w _X is the projection in the x direction, w _X ^T is the transposed projection, w _Y is the projection in the y direction, w _Y ^T is the transposed projection, the X matrix is the abscissa, X ^T is the transposed matrix, Y Matrix as abscissa, Y ^T is the transposed matrix, E[w _Y ^T XY ^T w _Y ] represents the expectation of w _X ^T XY ^T w _Y , E[w _X ^T XX ^T w _X ] represents w _X ^T XX ^T w The expectation of _X , E[w _Y ^T XY ^T w _Y ] represents the expectation of w _Y ^T XY ^T w _Y ;

According to the obtained maximum correlation coefficient and the second largest correlation coefficient of the RDA and QDA algorithms, the feature vector F is generated:

Here, it is only necessary to extract the correlation coefficients of the two methods, and then obtain a matrix in the following form:

是QDA的最大相关系数，

是QDA的第二大相关系数，

是RDA的最大相关系数，

是RDA的第二大相关系数；in,

is the maximum correlation coefficient of QDA,

is the second largest correlation coefficient of QDA,

is the maximum correlation coefficient of RDA,

is the second largest correlation coefficient of RDA;

S43. Divide the feature vector F into two categories: RDA-false and QDA-false through the projection classification unit

According to the classification results of the two algorithms, RDA and QDA, all training data trials are divided into four categories: both-true, both-false, RDA-false and QDA-false. In a double-correct trial, the two algorithms made the same and correct decision. In the double-sham test, the decisions of both methods were inconsistent with the participants' intentions. The proposed decision fusion method of choosing one of the two wrong decisions also gives a wrong result. So choose to train the decision fusion method with RDA-false and QDA-false trials (where only one decision in RDA and QDA is correct). In the soft boundary objective function, δ is a slack variable, which indicates whether the sample v _j is within the edge and needs to be adjusted. C is an adjustment coefficient that controls the trade-off between width and misclassification. The slack variable is used to determine whether the point is within the range, that is, only RDA is selected. -false and QDA-false categories.

The projection classification unit uses the linear SVM classifier to project within the range of the soft boundary objective function, and projects the feature vector F into a scalar value.

是线性核函数；Among them, v _j , (j=1,2,...,N) is the support vector, used to determine the maximum edge plane of the classifier, a _j > 0, is a parameter that can be changed and adjusted, y _j refers to the jth support the category of the vector, F is the feature vector, b is the bias,

is a linear kernel function;

The soft boundary objective function is:

Among them, δ is the slack variable, indicating whether the sample v _j is within the edge, the degree of adjustment is required, C is the adjustment coefficient that controls the trade-off between width and misclassification, and the slack variable is used to determine whether the point is within the range.

By solving the soft boundary objective function, the maximum value 2/||w|| is obtained, which is used as the boundary line. According to the position of the scalar value projection point, the features in the feature vector F are divided into RDA-false and QDA-false. class, one side of the boundary line is RDA-false class, and the other side is QDA-false class;

S44, select the decision output of the RDA or QDA algorithm according to the classification result by the decision selection unit;

As shown in Equation (20), RDA-false and QDA-false (where only one decision in RDA and QDA is correct) is chosen to train the decision fusion method, and if an RDA-false result is obtained, the module outputs a QDA decision. Otherwise, the module outputs the RDA decision.

Feature decision fusion unit is mainly composed of feature extraction unit, projection classification unit and decision selection unit. The feature decision fuser inputs the decisions and correlation coefficients of the RDA and QDA algorithms, selects and outputs more correct decisions.

The effect of the inventive method is compared and verified below:

During the testing and training phases of decision fusion, performance is estimated using a cross-validation where 5 datasets are selected for training decision fusion and 1 is selected for testing. During the training phase, we used another cross-validation to extract RDA-false and QDA-false features. Specifically, the RDA and QDA algorithms are trained using 4 blocks in the training dataset, and the remaining blocks are classified in each round. According to the classification results, the decision fusion feature F of RDA-false and QDA-false trials is extracted and recorded in 5 epochs of each training cycle, i.e. a total of 200 trials are tested. The decision fusion method is trained using recorded RDA-false and QDA-false features. In this method, we use the proposed decision fusion to synthesize five LDA-based classification algorithms: LDA, QDA, RDA, nearest mean, and weighted nearest mean. The classification accuracy of all combinations is estimated, and the classification accuracy and information transfer rate (ITR) of all combinations are estimated to judge the integrated performance.

Figure 2 illustrates the training and testing process of the proposed decision fusion method. Performance was estimated using a leave-one-out cross-validation, where 5 blocks of the dataset were chosen to train the decision fusion and 1 block was chosen to test. During the training phase, we used another cross-validation to extract RDA-false and QDA-false features. Specifically, the RDA and QDA algorithms are trained using 4 blocks in the training dataset, and the remaining blocks are classified in each round. According to the classification results, the decision fusion feature F of RDA-false and QDA-false trials is extracted and recorded in 5 epochs of each training cycle, i.e. a total of 200 trials are tested. The decision fusion method is trained using the recorded RDAfalse and QDA-false features.

In the testing phase, we synthesized 5 LDA-based classification algorithms using the proposed decision fusion method: LDA, QDA, RDA, nearest mean, and weighted nearest mean. The classification accuracy was estimated for all combinations. Classification accuracy and information transfer rate (ITR) were estimated for all combinations. The ITR in bits per minute is defined as follows:

Among them, P is the accuracy rate, N is the number of classes (that is, N=120 in this study), and T is the time required for one selection.

In terms of ensemble results, its performance is evaluated with a data length of 1 second. The resulting data consists of 5 rows and 5 columns, each row corresponding to an LDA-based algorithm. The main diagonal cells represent the average precision of each algorithm, and the other cells represent the average precision of the decision fusion method that integrates the two corresponding algorithms. For example, the accuracy of Decision Fusion-QDA&LDA method with data length of 1s is 90.56%, which is higher than 86.36% of LDA method, but lower than 93.70% of QDA method. However, there is a 7.34% difference in accuracy between QDA and LDA methods. Test results also show that the performance of decision fusion methods that combine QDA or RDA algorithms with low-precision algorithms degrades. In this study, the classification accuracy of the two algorithms before and after each decision fusion combination was calculated. These results show that the decision fusion method does not improve the overall classification accuracy when the accuracy of the two algorithms is very different. The decision fusion-QDA&RDA method, which combines two algorithms with similar performance, has the highest accuracy rate of 94.21% when the data length is 1s.

Comprehensive LDA, QDA, RDA, recent mean, weighted recent mean of five LDA-based classification algorithms, the classification accuracy of the proposed decision fusion method is shown in the following table:

Performance is evaluated with a data length of 1 second. The resulting data consists of 5 rows and 5 columns, each row corresponding to an LDA-based algorithm. The main diagonal cells represent the average precision of each algorithm, and the other cells represent the average precision of the decision fusion method that integrates the two corresponding algorithms.

In the results of the table, the accuracy of the decision fusion-QDA&LDA method with a data length of 1s is 90.56%, which is higher than the 86.36% of the LDA method, but lower than the 93.70% of the QDA method. However, there is a 7.34% difference in accuracy between QDA and LDA methods. The test results in Figure 3 also show that the decision fusion methods that combine QDA or RDA algorithms with low-precision algorithms degrade in performance. In this study, the classification accuracy of the two algorithms before and after each decision fusion combination was calculated. These results show that the fusion of the two algorithms does not improve the overall classification accuracy when the accuracy of the two algorithms is very different. The decision fusion-QDA&RDA method, which combines two algorithms with similar performance, has the highest accuracy rate of 94.21% when the data length is 1s.

The above-mentioned implementation cases are only to describe the preferred embodiments of the present invention, and do not limit the scope of the present invention. On the premise of not departing from the design spirit of the present invention, those of ordinary skill in the art can make various technical solutions of the present invention. Such deformations and improvements shall fall within the protection scope determined by the claims of the present invention.

Claims (4)

1. a kind of electroencephalogram consciousness dynamic classification method based on the feature decision fusion of linear analysis, is characterized in that, it comprises the following steps: S1. Collect the EEG signal data set X=(X1, X2, . . . , Xn) through the brain wave sensing helmet; S2, use regular discriminant analysis RDA to classify the signal data set X=(X1, X2, . . . , Xn) to obtain a correlation coefficient matrix ρ _n , which is defined as ρ _RDA ;

in,

is the matrix transpose of the weights, X is the data value,

is the mean of the data values; S3, use quadratic discriminant analysis _QDA to classify the signal data set X=( _X1 , X2, .

in,

is the matrix transpose of the weights, X is the data value,

is the mean of the data values; S4. Build a feature decision fusion device to perform feature integration and decision selection on the decisions and coefficients of RDA and QDA, and obtain dynamic classification of EEG consciousness. The specific steps are as follows; S41, constructing a feature decision fuser, and the feature decision fuser includes a feature extraction unit, a projection classification unit and a decision selection unit; S42, feature extraction is performed on the correlation coefficient of RDA and QDA by the feature extraction unit, and a feature vector F is generated; according to

The ratio judgment of , the larger the value, the larger the ρ, according to the expression:

Among them, w _X is the projection in the x direction, w _X ^T is the transposed projection, w _Y is the projection in the y direction, w _Y ^T is the transposed projection, the X matrix is the abscissa, X ^T is the transposed matrix, Y Matrix as abscissa, Y ^T is the transposed matrix, E[w _X ^T XY ^T w _Y ] represents the expectation of w _X ^T XY ^T w _Y , E[w _X ^T XX ^T w _X ] represents w _X ^T XX ^T w The expectation of _X , E[w _Y ^T XY ^T w _Y ] represents the expectation of w _Y ^T XY ^T w _Y ; Obtain the largest correlation coefficient and the second largest correlation coefficient of RDA and QDA, respectively; According to the obtained maximum correlation coefficient and the second largest correlation coefficient of the RDA and QDA algorithms, the feature vector F is generated:

in,

is the maximum correlation coefficient of QDA,

is the second largest correlation coefficient of QDA,

is the maximum correlation coefficient of RDA,

is the second largest correlation coefficient of RDA; S43. Divide the feature vector F into two categories: RDA-false and QDA-false through the projection classification unit; The projection classification unit uses the linear SVM classifier to project within the range of the soft boundary objective function, and projects the feature vector F into a scalar value.

Among them, v _j , j=1, 2, ..., N is the support vector, used to determine the maximum edge plane of the classifier, a _j > 0, is a parameter that can be changed and adjusted, y _j refers to the jth support vector category, F is the feature vector, b is the bias,

is a linear kernel function; The soft boundary objective function is:

Among them, δ _j is the slack variable, indicating whether the sample v _j is within the edge, the degree of adjustment is required, C is the adjustment coefficient that controls the trade-off between width and misclassification, the slack variable is used to determine whether the point is within the range, and W is the total number of samples; By solving the soft boundary objective function, the maximum value 2/||W|| is obtained, which is used as the boundary line. According to the position of the scalar value projection point, the features in the feature vector F are divided into RDA-false and QDA-false. kind; S44, select the decision output of the RDA or QDA algorithm according to the classification result by the decision selection unit; The decision selection unit is:

If an RDA-false result is obtained, the module outputs the QDA decision; otherwise, the module outputs the RDA decision to obtain a dynamic classification of EEG consciousness with better accuracy. 2 . The dynamic classification method for EEG consciousness based on feature decision fusion based on linear analysis according to claim 1 , wherein in step S1 , an EEG sensor helmet is used to collect EEG signals for Emotiv. 3 . 3. the electroencephalographic consciousness dynamic classification method of the feature decision fusion based on linear analysis according to claim 1, is characterized in that, the solving step of ρ _RDA in step S2 is specifically as follows: The dataset X = (X1, X2, ..., Xn) is assigned to one of the K groups of classes. In the training data, the class of the data is known, so the prior probability and mean of class k are:

in,

is the prior probability of class k,

is the mean of class k, ω is the total number of samples, ω _k is the number of samples of class k, X _n is the sample point,

is the mean of the values of category k; Regular discriminant analysis RDA improves the effects of multicollinearity by modifying singular covariance values; the sample covariance for each class is estimated as follows:

where ω is the total number of samples, X _n is the sample points,

is the mean of category k; The covariance matrix is further adjusted by introducing a shrinkage parameter γ:

where λ is the regularization parameter, 0≤λ≤1, p is the dimension of the independent variable, I is the identity matrix, and γ is the shrinkage parameter; The optimization objective is J(w):

The above formula is the generalized Rayleigh quotient of _Sk and S, where _Sk = ω _k ∑ _k ,

This is the goal that QDA wants to maximize. The maximum value of J is the maximum eigenvalue of the matrix S _k ^-1 S, and the corresponding eigenvector is the eigenvector corresponding to the maximum eigenvalue of S _k ^-1 S; solve w=S _k ^{- 1} (μ ₁ -μ ₂ ) is the determined optimal projection direction, w ^T is the transposed projection, and the samples in the training set are projected to the w direction to obtain: y=w ^T X (6)

in,

is the matrix transpose of the weights, X is the data value,

is the mean of the data values. 4. the electroencephalogram consciousness dynamic classification method of the feature decision fusion based on linear analysis according to claim 1, is characterized in that, the solving step of ρ _QDA in step S3 is specifically as follows: Assuming that the sample data set X = (X1, X2, ..., Xn) obeys the multivariate Gaussian distribution, μ _i is the mean vector, expressed as:

Compute the sample covariance matrix ∑j:

Among them, j=1, 2; The intra-class divergence matrix S _w can be obtained as:

At the same time, the inter-class divergence matrix S _b is defined as: S _b =(μ ₁ -μ ₂ )(μ ₁ -μ ₂ ) ^T (12) The optimization objective is J(w):

The above formula is the generalized Rayleigh quotient of S _w and S _b , which is the goal of QDA to maximize. The maximum value of J is the largest eigenvalue of the matrix S _w ^-1 S _b , and the corresponding is S _w ^-1 S _b The eigenvector corresponding to the largest eigenvalue of ; obtain w=S _w ^-1 (μ ₁ -μ ₂ ), which is the determined optimal projection direction, and project the samples in the training set to the w direction to obtain: y=w ^T X (14)

Images