CN105930864A - EEG (electroencephalogram) signal feature classification method based on ABC-SVM - Google Patents

Abstract

The invention relates to an EEG (electroencephalogram) signal feature classification method based on an ABC-SVM. The method comprises the steps: firstly carrying out EEG signal feature extraction through employing a CSP algorithm; secondly carrying out the optimization of a penalty factor C and a kernel parameter g of a SVM (support vector machine) through employing an artificial bee colony algorithm; finally carrying out the training of an SVM classifier through the obtained optimal parameter, and carrying out the classification prediction of samples through employing the trained classifier. Compared with an SVM classification recognition method optimized through a conventional algorithm, the method can effectively improve the classification recognition rate of an EEG signal, and is remarkably superior to a conventional classification recognition method.

Description

A kind of EEG signals tagsort method based on ABC-SVM

Technical field

The present invention relates to the feature extraction and classifying method of EEG signals, particularly to EEG signals tagsort method based on ABC-SVM.

Background technology

Brain-computer interface technology (Brain-computer Interface, BCI) being a kind of new technique having merged multiple subjects such as artificial intelligence, computer science and information, biomedical engineering and Neuscience, this technology suffers from quite varied application prospect in medical treatment, disability rehabilitation and brain cognitive science field.The feature of BCI technology is neural by human peripheral, muscle and the normal physiological path of bone, and allows brain and external equipment (computer or relevant instrument etc.) directly transmit information or control command.

The pattern classification of EEG signals is a ring particularly important in BCI system, and the quality of classification performance is related to the real-time of BCI system, availability and stability.So, numerous algorithms select suitable method and is improved, exploring new method particularly important with the accuracy rate improving classification simultaneously.SVMs (Support vector machine, SVM) it is the novel universal algorithm that statistical theory and Structural risk minization principle are combined and grown up, between complexity (i.e. the study precision to specific training sample) and the learning ability (identifying the ability of arbitrary sample the most error-free) of model, best compromise is sought according to limited sample information, solving small sample, good Generalization Capability is had in the non-linear and identification problem of High-Dimensional Model, thus in pattern-recognition, image procossing, time series forecasting, the fields such as fault diagnosis have obtained increasingly being widely applied.There is an outstanding problem in SVM in concrete application process, the most how to select to affect the key model parameter of algorithm performance, because the selection of parameter determines learning performance and the generalization ability of SVM.The most conventional Model Parameter Optimization method mainly has gradient descent method, genetic algorithm and particle cluster algorithm etc., but these optimization methods have respective limitation.Such as gradient descent algorithm, stability is the highest and inefficient；Genetic algorithm convergence rate is relatively slow and is easily trapped into local optimum；Particle swarm optimization algorithm is compared to genetic algorithm, although principle is simple, easily realizes, fast convergence rate, but the later stage is easily absorbed in local optimum, it is impossible to reach optimal classification effect.Artificial bee colony algorithm (Artificial Bee Colony, ABC) is to be inspired in a kind of swarm intelligence algorithm of bee colony cooperation gathering honey behavior.The biological foundation of this algorithm is: employ honeybee, observes honeybee and search bee completes the gathering honey under different natural environments by the way of information interchange and role transforming, by the collection of food source information with shared, find the optimal solution of problem.Karaboga etc. are proved by a large amount of Benchmark function test experiments, this algorithm is by the cooperation between different work post honeybees, solve extension new explanation territory and carry out the contradiction between fine search in known solution territory, largely avoid and be absorbed in locally optimal solution problem, possess optimization performance more more preferable than traditional optimization.

Summary of the invention

It is an object of the invention to the good characteristic utilizing artificial bee colony algorithm to be shown as a kind of new colony intelligence optimized algorithm, kernel functional parameter and penalty factor to SVMs are iterated optimizing, it is proposed that EEG signals tagsort method based on ABC-SVM.

The purpose of the present invention can be achieved through the following technical solutions:

EEG signals tagsort method based on ABC-SVM, the method comprises the following steps:

Step 1. utilizes CSP algorithm that EEG signals are carried out feature extraction, obtains sampling feature vectors f_i；

Step 2. utilizes artificial bee colony algorithm that kernel functional parameter and the penalty factor of SVMs are iterated optimizing；

SVM classifier is trained by the optimized parameter after step 3. utilizes artificial bee colony algorithm optimization, utilizes the grader trained that sample carries out classification prediction.

Wherein in step 1, EEG feature extraction obtains characteristic vector and specifically comprises the following steps that

Using CSP algorithm that EEG signals are carried out feature extraction, if the channel number of experimental data is N, the sampling number of each passage is T, and the EEG data once tested is X_n[N*T], n is the classification of EEG signals.First, the covariance matrix of training sample is solved；Secondly, covariance matrix is decomposed；Then, construct spatial filter, maximum for reaching two class signal differences, choose front k Special composition wave filter together with rear k characteristic vector, primary signal is projected to this wave filter and can obtain signal Z newly_i.Finally, k is calculated to new signal Z_iVariance, and it is taken the logarithm and carries out standardized operation and obtain feature:

$f_i=\frac{log\left(\mathrm{var}\right(Z_i\left)\right)}{\underset{j=1}{\overset{2k}{\Σ}}log\left(\mathrm{var}\right(Z_j\left)\right)}$

In formula, f_iFor the EEG signals characteristic vector extracted, var (Z_i) it is Z_iVariance.

As preferably, kernel function K (x, the f in step 2_i) select Radial basis kernel function (RBF), formula is as follows:

K(x,f_i)=exp (-| x-f_i|²)/g²

Wherein () is inner product, x, f_i∈Rⁿ, f_iBeing characterized vector, g is nuclear parameter, then the optimal decision function formula of SVMs is converted to:

$f\left(x\right)=sign\[\underset{i=1}{\overset{l}{\Σ}}{a}_i^{*}{y}_{i}K(x,f_i)+b^{*}\],0\≤a_i\≤C$

In formula, C is penalty factor, a_iFor corresponding Lagrange coefficient, b^*For classification thresholds.

As preferably, the parameter to SVMs of the artificial bee colony algorithm described in step 2, i.e. nuclear parameter g and penalty factor, it is iterated specifically comprising the following steps that of optimizing

1) initializing: initializing bee colony size is CS, object function maximum assessment number of times is MCN, and the most very much not update times is Limit, number of parameters to be optimized is Dim, and the bound of parameter to be optimized is respectively ub, lb, then solution space size is CS, and the number employing honeybee and observation honeybee is CS/2.Initially dissolve for:

x_ij=lb+rand (ub-lb)

Wherein, rand represents the uniformly random distribution that scope is [0,1], i ∈ 1,2 ..., CS}, j ∈ 1,2 ..., Dim}.

2) honeybee is employed to search new explanation: to employ honeybee in initial solution x_ijNeighborhood produces new solution v_ij:

Wherein, i ≠ k,I.e. scope is the uniformly random distribution of [-1,1].Calculated the fitness of initial solution and new explanation by object function and fitness function, if the fitness of new explanation is higher than initial solution, then replaces initial solution with new explanation, do not replace.

3) observation honeybee selects to solve: the probability that observation honeybee solves according to the solution employing honeybee to transmit and the calculating selection of fitness information:

$p_i=\frac{{\mathrm{fit}}_i}{\underset{i}{\overset{CS}{\Σ}}{\mathrm{fit}}_i}$

Wherein, fit_iRepresent the fitness of the i-th solution.Observe honeybee to select to solve according to the mode of roulette, then observe and produce new explanation around honeybee neighborhood around selected solution, calculate and select to solve and after the fitness of new explanation, select to solve between new explanation and former selection solution according to greedy algorithm.

4) search bee occurs: if solution is after Limit time circulates, the quality of solution is not improved, then employing honeybee changing role is search bee, abandons this solution and randomly generates new explanation, and being set to 0 by iterations.

5) terminate algorithm: judge whether iterations reaches MCN, if not up to, then algorithm is normally carried out, otherwise terminate algorithm, export optimal solution.

Beneficial effects of the present invention: after utilizing CSP to carry out feature extraction, will based on ABC-SVM Classification and Identification result with use single SVM carries out Classification and Identification, PSO-SVM and GA-SVM Classification and Identification result contrasts, result shows, the accuracy that EEG signals are classified by the SVM classifier after using ABC to optimize is higher, effectively raises Classification and Identification rate.

Accompanying drawing explanation

Fig. 1 is that certain tests ABC optimizing fitness curve map；

Fig. 2 is that certain tests ABC-SVM classification accuracy schematic diagram；

Fig. 3 is the accuracy rate statistical chart of 100 experiments of four experimenters.

Detailed description of the invention

Below in conjunction with accompanying drawing, the invention will be further described.

The present invention comprises the following steps:

Step 1. utilizes CSP algorithm that EEG signals are carried out feature extraction, obtains sampling feature vectors f_i；

Step 2. utilizes artificial bee colony algorithm that kernel functional parameter and the penalty factor of SVMs are iterated optimizing；

SVM classifier is trained by the optimized parameter after step 3. utilizes artificial bee colony algorithm optimization, utilizes the grader trained that sample carries out classification prediction.

Wherein in step 1, EEG feature extraction obtains characteristic vector and specifically comprises the following steps that

Using CSP algorithm that EEG signals are carried out feature extraction, if the channel number of experimental data is N, the sampling number of each passage is T, and the EEG data once tested is X_n[N*T], n is the classification of EEG signals.First, the covariance matrix of training sample is solved；Secondly, covariance matrix is decomposed；Then, construct spatial filter, maximum for reaching two class signal differences, choose front k Special composition wave filter together with rear k characteristic vector, primary signal is projected to this wave filter and can obtain signal Z newly_i.Finally, k is calculated to new signal Z_iVariance, and it is taken the logarithm and carries out standardized operation and obtain feature:

$f_i=\frac{log\left(\mathrm{var}\right(Z_i\left)\right)}{\underset{j=1}{\overset{2k}{\Σ}}log\left(\mathrm{var}\right(Z_j\left)\right)}$

In formula, f_iFor the EEG signals characteristic vector extracted, var (Z_i) it is Z_iVariance.

As preferably, kernel function K (x, the f in step 2_i) select Radial basis kernel function (RBF), formula is as follows:

K(x,f_i)=exp (-| x-f_i|²)/g²

Wherein () is inner product, x, f_i∈Rⁿ, f_iBeing characterized vector, g is nuclear parameter, then the optimal decision function formula of SVMs is converted to:

$f\left(x\right)=sign\[\underset{i=1}{\overset{l}{\Σ}}{a}_i^{*}{y}_{i}K(x,f_i)+b^{*}\],0\≤a_i\≤C$

In formula, C is penalty factor, a_iFor corresponding Lagrange coefficient, b^*For classification thresholds.

As preferably, the parameter to SVMs of the artificial bee colony algorithm described in step 2, i.e. nuclear parameter g and penalty factor, it is iterated specifically comprising the following steps that of optimizing

1) initializing: initializing bee colony size is CS, object function maximum assessment number of times is MCN, and the most very much not update times is Limit, number of parameters to be optimized is Dim, and the bound of parameter to be optimized is respectively ub, lb, then solution space size is CS, and the number employing honeybee and observation honeybee is CS/2.Initially dissolve for:

x_ij=lb+rand (ub-lb)

Wherein, rand represents the uniformly random distribution that scope is [0,1], i ∈ 1,2 ..., CS}, j ∈ 1,2 ..., Dim}.

2) honeybee is employed to search new explanation: to employ honeybee in initial solution x_ijNeighborhood produces new solution v_ij:

Wherein, i ≠ k,I.e. scope is the uniformly random distribution of [-1,1].Calculated the fitness of initial solution and new explanation by object function and fitness function, if the fitness of new explanation is higher than initial solution, then replaces initial solution with new explanation, do not replace.

3) observation honeybee selects to solve: the probability that observation honeybee solves according to the solution employing honeybee to transmit and the calculating selection of fitness information:

$p_i=\frac{{\mathrm{fit}}_i}{\underset{i}{\overset{CS}{\Σ}}{\mathrm{fit}}_i}$

Wherein, fit_iRepresent the fitness of the i-th solution.Observe honeybee to select to solve according to the mode of roulette, then observe and produce new explanation around honeybee neighborhood around selected solution, calculate and select to solve and after the fitness of new explanation, select to solve between new explanation and former selection solution according to greedy algorithm.

4) search bee occurs: if solution is after Limit time circulates, the quality of solution is not improved, then employing honeybee changing role is search bee, abandons this solution and randomly generates new explanation, and being set to 0 by iterations.

5) terminate algorithm: judge whether iterations reaches MCN, if not up to, then algorithm is normally carried out, otherwise terminate algorithm, export optimal solution.

Certain from Fig. 1 is once tested ABC optimizing fitness curve map and be can be seen that, along with the increase of iterations, accuracy rate gradually tends to optimal adaptation degree, when meeting stopping criterion for iteration, output optimal value of the parameter c=19.1897, g=0.64, for support vector cassification fast searching to optimal penalty parameter c and nuclear parameter g, and ABC is parallel global search based on population, regulation parameter is few, fast convergence rate, it is to avoid be absorbed in local optimum, embodies the clear superiority utilizing ABC-SVM.

Fig. 2 is that certain tests ABC-SVM classification accuracy schematic diagram, randomly selects the label of 120 groups of samples in 200 groups of samples and correspondence as training set, and training set, as test set, is trained by remaining with SVM.The when of doing classification prediction, by using optimal models as input parameter, selected test set and penalty factor and nuclear parameter g etc., it was predicted that go out the tag along sort of test set data.Correct label is compared with prediction label, draws classification accuracy rate percentage.The wherein red type representing prediction test set label, the blue type representing actual test set label, can significantly be seen by Fig. 2, in 80 groups of experiments, most experiment prediction label and physical tags coincide, the label only having indivedual test sets is not squared with physical tags, and classification accuracy reaches 91.25%, thus demonstrates the validity of algorithm.

Fig. 3 is that tetra-experimenters of b, f, g test, and repeat respectively to test 100 times by 4 group data sets to a, in order to make experimental result more very clear, the classification accuracy statistics that 100 experiments draw is drawn in the way of box traction substation, as shown in Figure 3.Wherein, abscissa represents four experimenters, and ordinate represents recognition accuracy, and the two ends of box traction substation represent four mark up and down of 100 experiment accuracys rate respectively, and center line represents the median of accuracy rate, "+" represent gentle abnormity point.Through observation shows that, in addition to the accuracy rate of experimenter b is lower slightly, the neutrality line of other three experimenters, all more than 0.85, illustrates that their Average Accuracy is all higher than 85%, and high-accuracy is close to 93%.For all experimenters, all there is an other gentle abnormity point.

In order to contrast sorting technique and other sorting technique optimizing SVM based on artificial bee colony, the experimental result that ABC-SVM with SVM, the SVM of genetic algorithm optimization and particle cluster algorithm optimize SVM is compared, for experimenter a, b, f, g provide nominal data, 4 kinds of sorting techniques repeat to test 100 times respectively, and statistical average accuracy rate and standard deviation as shown in table 1.

1 four kinds of algorithm classification accuracys rate of table (means standard deviation (%))

ABC-SVM classification results is carried out contrast as shown in table 1 with single SVM, GA-SVM and PSO-SVM.As can be seen from the table: for a, tetra-experimenters of b, f, g, different sorting techniques is different for the classification performance quality of different experimenters.The classification accuracy data of the different sorting technique of contrast, from the point of view of the average classification accuracy of some experimenter, compared to other sorting technique, ABC-SVM obtains more preferable classifying quality, and the highest average classification accuracy can reach 90.25%.Simultaneously for the average classification accuracy that four experimenters are total, the classification accuracy of ABC-SVM is the highest, has reached 85.49%, and the average classification results of PSO-SVM, GA-SVM and SVM is respectively 83.61%, 83.48% and 82.14%.Additionally, it is no matter the SVM that optimizes of traditional algorithm or the svm classifier method that artificial bee colony algorithm optimizes, the classification accuracy higher than former SVM algorithm can be obtained, again demonstrate the feasibility of ABC-SVM algorithm, show that ABC-SVM shows obvious advantage at eeg signal classification, it is possible to effective raising eeg signal classification accuracy rate.

Claims (4)

1. an EEG signals tagsort method based on ABC-SVM, it is characterised in that the party Method comprises the following steps:

Step 1. utilizes CSP algorithm that EEG signals are carried out feature extraction, obtain sample characteristics to Amount f_i；

Step 2. utilizes artificial bee colony algorithm to enter kernel functional parameter and the penalty factor of SVMs Row iteration optimizing；

SVM classifier is trained by the optimized parameter after step 3. utilizes ABC to optimize, and utilizes The grader trained carries out classification prediction to sample.

A kind of EEG signals tagsort based on ABC-SVM the most according to claim 1 Method, it is characterised in that: wherein in step 1 EEG feature extraction to obtain characteristic vector concrete Step is as follows:

Use CSP algorithm that EEG signals are carried out feature extraction, if the channel number of experimental data is N, The sampling number of each passage is T, and the EEG data once tested is X_n[N*T], n is the class of EEG signals Not；First, the covariance matrix of training sample is solved；Secondly, covariance matrix is decomposed；Then, structure Make spatial filter, maximum for reaching two class signal differences, choose front k together with rear k characteristic vector Special composition wave filter, projects to primary signal this wave filter and can obtain signal Z newly_i；Finally, k is calculated To new signal Z_iVariance, and it is taken the logarithm and carries out standardized operation and obtain feature:

$f_i=\frac{log\left(var\right(Z_i\left)\right)}{\underset{j=1}{\overset{2k}{\Σ}}log\left(var\right(Z_j\left)\right)}$

In formula, f_iFor the EEG signals characteristic vector extracted, var (Z_i) it is Z_iVariance.

A kind of EEG signals tagsort method based on ABC-SVM the most according to claim 1, It is characterized in that: kernel function K (x, the f in step 2_i) select Radial basis kernel function, formula is as follows:

K(x,f_i)=exp (-| x-f_i|²)/g²

Wherein () is inner product, x, f_i∈Rⁿ, f_iBeing characterized vector, g is nuclear parameter, then the optimum of SVMs is certainly Plan function formula is converted to:

$f\left(x\right)=sign\[\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{a}_i^{*}{y}_{i}K(x,f_i)+b^{*}\],0\≤a_i\≤C$

In formula, C is penalty factor, a_iFor corresponding Lagrange coefficient, b^*For classification thresholds.

A kind of EEG signals tagsort based on ABC-SVM the most according to claim 1 Method, it is characterised in that: the parameter to SVMs of the artificial bee colony algorithm described in step 2, I.e. nuclear parameter g and penalty factor, is iterated specifically comprising the following steps that of optimizing

1) initializing: initializing bee colony size is CS, object function maximum assessment number of times is MCN, the most very much not update times is Limit, and number of parameters to be optimized is Dim, parameter to be optimized Bound be respectively ub, lb, then solution space size is CS, employ honeybee and observe honeybee number For CS/2；Initially dissolve for:

x_ij=lb+rand (ub-lb)

Wherein, rand represents the uniformly random distribution that scope is [0,1], i ∈ 1,2 ..., CS}, j∈{1,2,…,Dim}；

2) honeybee is employed to search new explanation: to employ honeybee in initial solution x_ijNeighborhood produces new solution v_ij:

Wherein, i ≠ k,I.e. scope is the uniformly random distribution of [-1,1]；Pass through mesh Scalar functions and fitness function calculate the fitness of initial solution and new explanation, if the fitness of new explanation is higher than Initial solution, then replace initial solution with new explanation, do not replace；

3) observe honeybee to select to solve: observe honeybee according to the solution employing honeybee to transmit and fitness information meter The probability that calculation selection solves:

$p_i=\frac{{\mathrm{fit}}_i}{\underset{i}{\overset{CS}{\Σ}}{\mathrm{fit}}_i}$

Wherein, fit_iRepresent the fitness of the i-th solution；Observe honeybee to select to solve according to the mode of roulette, Then observe and produce new explanation around honeybee neighborhood around selected solution, calculate and select to solve and new explanation After fitness, select to solve between new explanation and former selection solution according to greedy algorithm；

4) search bee occurs: if solution is after Limit time circulates, the quality of solution is not improved, Then employing honeybee changing role is search bee, abandons this solution and randomly generates new explanation, and by iteration time Number sets to 0；

5) algorithm is terminated: judge whether iterations reaches MCN, if not up to, then algorithm normally enters OK, otherwise terminate algorithm, export optimal solution.