CN105677043A - Two-stage self-adaptive training method for motor imagery brain-computer interface - Google Patents

Abstract

The invention relates to a two-stage self-adaptive training method of a motor imagery brain-computer interface, which includes a single trust stage experienced first and a mutual trust stage entered again. Single trust means that the system trusts the data and updates the classifier with the data. At this stage, the system first conducts preliminary training based on multiple trials conducted by users to obtain a preliminary feasible classifier, and then uses incremental learning to continuously update the classifier. Mutual trust means that the system trusts the data and the classifier at the same time, so that the two can adapt to each other. At this stage, the system first uses the classifier obtained in the single-trust stage to identify. After the user conducts a certain amount of trials, according to the feedback results, the data is optimized by using SVM to find support vectors, and the classification is continuously updated by incremental learning. The classifier is used to identify and feed back the subsequent trial with a new classifier, and this process is repeated until the end of the training. The method provided by the invention can enhance the mutual adaptation between the classifier and the user, and has short time consumption and high accuracy.

Description

A two-stage adaptive training method for motor imagery brain-computer interface

technical field

The invention relates to the mutual adaptation problem of the human-machine training process of the brain-computer interface, and is an adaptive training method for the brain-computer interface of the motor imagery type.

Background technique

There is a problem of mutual adaptation between the user of the BCI and the BCI system. In the process of man-machine training, the brain-computer interface system continuously collects data and adapts to the characteristics of the user through machine learning, and the user also actively adjusts the activities in the brain to adapt to the working method of the brain-computer interface by observing the feedback of the brain-computer interface system . The mutual adaptation between the BCI user and the BCI system is a dynamic process. Collecting training data without considering user feedback, and only carrying out machine learning on a static training set is inconsistent with this characteristic of the brain-computer interface, which is not conducive to improving the performance of the brain-computer interface.

Contents of the invention

In view of this, the object of the present invention is to provide a training method for man-machine mutual adaptation for motor imagery brain-computer interface. The invention divides the man-machine training process of the motor imagery brain-computer interface into two stages, the former stage is a single-trust stage, and the latter stage is a mutual-trust stage. In the single-trust stage, the brain-computer interface system trusts the data and uses incremental learning to continuously update the classifier with new data. In the mutual trust stage, the brain-computer interface system trusts both the data and the classifier, allowing the two to adapt to each other.

The present invention is implemented in the following manner: a two-stage adaptive training method for motor imagery brain-computer interface, comprising the following steps:

Step S1: This step is in the single-trust stage. The motor imagery brain-computer interface system trusts the data and uses the data to update the classifier; initially, the user conducts multiple trials of motor imagery, the system collects samples online, and uses the LDA/QR algorithm to obtain a preliminary transfer matrix G, and calculate the classification center vector set C to form an initial classifier; the user continues to carry out motor imagery, and the system uses the classifier to identify the user’s motor imagery online and give feedback to the user, and obtain new samples at the same time; each time a trial is completed, The system uses the ILDA/QR algorithm to update the transfer matrix G, and calculates the classification center vector set C to form a new classifier and use it for the identification and feedback of the next trial until the end of this stage;

Step S2: This step is in the stage of mutual trust. The motor imagery brain-computer interface system trusts both the data and the classifier at the same time, so that the two can adapt to each other. Initially, the system uses the classifier finally obtained in step S1 to identify the user's motor imagery online and report to the user Feedback: Whenever the user conducts a certain number of trials, the system uses the mutual trust optimization method to screen new samples, uses the LDA/QR algorithm or ILDA/QR algorithm to update the transfer matrix G, and calculates the classification center vector set C to form a new classifier and Use it for subsequent recognition and feedback until the end of man-machine training.

Further, the step S1 specifically includes the following steps:

Step S11: The user carries out multiple trials of motor imagery, the system collects signals online, intercepts a section of signal every other time interval, and converts it into an m-dimensional feature vector through feature extraction, which is recorded as x;

Step S12: use all the eigenvectors obtained in the step S11 to construct a data matrix A, and construct a matrix E according to the number of categories k;

Step S13: Execute the LDA/QR algorithm to obtain the optimal transition matrix G, and calculate the class center vector set C to form an initial classifier;

Step S14: The user carries out a trial motor imagery; the system collects EEG signals online, intercepts a segment of the signal every other time interval, converts it into a feature vector x through feature extraction, identifies x with a classifier, and moves the controlled object to the user according to the recognition result feedback;

Step S15: Execute the ILDA/QR algorithm, update the transition matrix G, then update the data matrix A, and calculate the class center vector set C to form a new classifier;

Step S16: This stage ends; otherwise, use the new classifier for the identification and feedback of the next trial, and return to step S14.

Further, the step S2 specifically includes the following steps:

Step S21: it is 0 to set sign flag, and the construction container is used for depositing k class motor imagery sample;

Step S22: The user carries out a trial motor imagery; the system collects EEG signals online, intercepts a segment of the signal every other time interval, converts it into a feature vector x through feature extraction, identifies x with a classifier, and moves the controlled object to the user according to the recognition result feedback;

Step S23: when the current trial ends, if the controlled object hits the target, store all samples of the trial into the container, and record the time it takes for the trial to hit the target;

Step S24: If the number of trials does not meet the adaptive update condition, return to step S22; otherwise, the system uses the mutual trust optimization method to screen new samples;

Step S25: If the flag is equal to 0, construct a data matrix A, construct a matrix E according to the number of categories k, execute the LDA/QR algorithm to obtain a new transition matrix G, and set the flag to 1; otherwise, execute the ILDA/QR algorithm to update the transition matrix G , and update the data matrix A;

Step S26: Calculate a new class center vector set C according to the transition matrix G and the data matrix A, form a new classifier, and empty the container;

Step S27: the man-machine training is over; otherwise, return to step S22.

Further, the mutual trust optimization method described in step S24 includes the following steps:

Step S41: Filter trials according to the time it takes to hit the target; use the time-consuming corresponding to all trials in the container to form a preliminary set, find the minimum time-consuming, and other time-consuming times are generated at their symmetrical positions with the minimum time-consuming time as the symmetric axis A virtual time, all the virtual time is added to the preliminary set to form a reference set, the standard deviation of the reference set is calculated, and the trial that takes less time than the sum of the minimum time spent and the standard deviation is selected;

Step S42: Use the support vector method to screen samples; all samples corresponding to the trial selected in step S41 are added to the candidate sample set, train on the candidate sample set with SVM, and select those samples that are determined to be support vectors;

Wherein, the trial represents an experimental unit in the human-computer training process of the motor imagery brain-computer interface; when the trial starts, the brain-computer interface system randomly sets a target; The system determines the direction to actually move the controlled object by identifying the type of motor imagery; the trial ends when the controlled object hits the target or times out.

Further, the method for constructing or updating the data matrix A described in steps S12 and S25 is:

Data matrix A=[x ₁ ,x ₂ ,…,x _n ]=[A ₁ ,…,A _k ]∈R ^m×n , x _i ∈R ^m×1 i=1,…,n represents an m-dimensional There are a total of n sample points in A; m is determined by the feature extraction method and is fixed during system operation; n is dynamically increased during system operation; each sample block matrix Represents the set of all sample points of the i-th class, and there are k classes in total, n _i i=1,...,k represents the number of samples of the i-th class,

Further, the specific method of constructing the matrix E according to the number of categories k described in steps S12 and S25 is:

matrix in $e_i={\left[\begin{array}{cccc}1& 1& ...& 1\end{array}\right]}^T\∈R^{{\mathrm{no}}_i},i=1,...,k.$

Further, the input of the LDA/QR algorithm is a data matrix A∈R ^m×n and E∈R ^n×k , and the output is a transfer matrix G∈R ^m×k ;

Specifically include the following steps:

Step S31: Calculate the economic QR decomposition of A, A=QR, Q∈R ^m×n , R∈R ^n×n , where Q is column-orthogonal and R is non-singular;

Step S32: Solve a lower triangular linear system R ^T H = E, and obtain H;

Step S33: Calculate G, G=QH.

Further, the input of the ILDA/QR algorithm is a column orthogonal matrix Q, an optimal transfer matrix G and a new sample x and its category l, and the output is an updated column orthogonal matrix Q and transfer matrix G;

Specifically include the following steps:

Step S41: Calculate r=Q ^T x, Q=[Q(x-Qr)/α];

Step S42: Calculate r=-G ^T x , r(l)=r(l)+1, r=r/α, G=G+Q(:,n+1)r ^T .

Further, the calculation method of the classification center vector set C is as follows: the data matrix A and the transition matrix G of the existing known categories are calculated P= ^AT G, $P=\left[\begin{array}{c}{p}_1\\ p_2\\ .\\ .\\ .\\ p_{\mathrm{no}}\end{array}\right]=\left[\begin{array}{c}{P}_1\\ P_2\\ .\\ .\\ .\\ P_{\mathrm{no}}\end{array}\right]\∈R^{\mathrm{no}\×k},$ Where p _j ∈ R ^1×k j=1,...,n represents the projection result corresponding to each sample point in the data matrix A, P _i ∈ R ^1×k i=1,...,k represents the i-th class The set of all p _j of ; calculate the center C _i ∈ R ^1×k of each class, $C_i=\frac{1}{{\mathrm{no}}_i}\underset{p_j\∈P_i}{\Σ}{p}_j,$ Get the class center vector set as $C=\left[\begin{array}{c}{C}_1\\ .\\ .\\ .\\ C_k\end{array}\right].$

Further, the specific classification method of the classifier is: the sample to be determined is x, calculate d _i =||x ^T GC _i || ₂ i=1,...,k, select the smallest d _i , and determine x for the corresponding category.

Before using the brain-computer interface system, users of the brain-computer interface need to go through a human-computer training process to make the two adapt to each other. In this process, the user's state is not static, and the user constantly adjusts his state to adapt to the working method of the brain-computer interface while observing the feedback of the brain-computer interface system. The quality of the data set collected by the brain-computer interface system changes with the change of the user's state. It is very important to enable the system to automatically select high-quality data during the human-machine training process. Therefore, compared with the prior art, the present invention has the following advantages:

1. The present invention enables the brain-computer interface system to automatically select samples during the human-machine training process, so that the good state of the brain-computer interface user can be highlighted in the new sample set.

2. The present invention adopts the method of incremental learning. When a new sample appears, the brain-computer interface system can quickly absorb the information contained in the new sample without retraining.

3. The present invention adopts a simple and efficient algorithm. The brain-computer interface system can update itself online during the human-computer training process, so that the mutual adaptation between the brain-computer interface system and the user can be dynamically completed during the human-computer training process.

In summary, the method provided by the present invention can enhance the mutual adaptability between the brain-computer interface system and the user during the man-machine training process.

Description of drawings

Fig. 1 is a schematic diagram of the overall process of the present invention.

FIG. 2 is a schematic flow chart of step S1 of the present invention.

Fig. 3 is a schematic flow chart of step S2 of the present invention.

Fig. 4 is a schematic diagram of a trial of motor imagery in the present invention.

detailed description

The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

This implementation provides a two-stage adaptive training method for motor imagery brain-computer interface, as shown in Figure 1, including the following steps:

Step S1: This step is in the single-trust stage. The motor imagery brain-computer interface system trusts the data and uses the data to update the classifier; initially, the user conducts multiple trials of motor imagery, the system collects samples online, and uses the LDA/QR algorithm to obtain a preliminary transfer matrix G, and calculate the classification center vector set C to form an initial classifier; the user continues to carry out motor imagery, and the system uses the classifier to identify the user’s motor imagery online and give feedback to the user, and obtain new samples at the same time; each time a trial is completed, The system uses the ILDA/QR algorithm to update the transfer matrix G, and calculates the classification center vector set C to form a new classifier and use it for the identification and feedback of the next trial until the end of this stage;

Step S2: This step is in the stage of mutual trust. The motor imagery brain-computer interface system trusts both the data and the classifier at the same time, so that the two can adapt to each other. Initially, the system uses the classifier finally obtained in step S1 to identify the user's motor imagery online and report to the user Feedback: Whenever the user conducts a certain number of trials, the system uses the mutual trust optimization method to screen new samples, uses the LDA/QR algorithm or ILDA/QR algorithm to update the transfer matrix G, and calculates the classification center vector set C to form a new classifier and Use it for subsequent recognition and feedback until the end of man-machine training.

In this embodiment, as shown in FIG. 2, the step S1 specifically includes the following steps:

Step S11: The user carries out multiple trials of motor imagery, the system collects signals online, intercepts a section of signal every other time interval, and converts it into an m-dimensional feature vector through feature extraction, which is recorded as x;

Step S12: use all the eigenvectors obtained in the step S11 to construct a data matrix A, and construct a matrix E according to the number of categories k;

Step S13: Execute the LDA/QR algorithm to obtain the optimal transition matrix G, and calculate the class center vector set C to form an initial classifier;

Step S14: The user carries out a trial motor imagery; the system collects EEG signals online, intercepts a segment of the signal every other time interval, converts it into a feature vector x through feature extraction, identifies x with a classifier, and moves the controlled object to the user according to the recognition result feedback;

Step S15: Execute the ILDA/QR algorithm, update the transition matrix G, then update the data matrix A, and calculate the class center vector set C to form a new classifier;

Step S16: This stage ends; otherwise, use the new classifier for the identification and feedback of the next trial, and return to step S14.

In this embodiment, as shown in FIG. 3, the step S2 specifically includes the following steps:

Step S21: it is 0 to set sign flag, and the construction container is used for depositing k class motor imagery sample;

Step S22: The user carries out a trial motor imagery; the system collects EEG signals online, intercepts a segment of the signal every other time interval, converts it into a feature vector x through feature extraction, identifies x with a classifier, and moves the controlled object to the user according to the recognition result feedback;

Step S23: when the current trial ends, if the controlled object hits the target, store all samples of the trial into the container, and record the time it takes for the trial to hit the target;

Step S24: If the number of trials does not meet the adaptive update condition, return to step S22; otherwise, the system uses the mutual trust optimization method to screen new samples;

Step S25: If the flag is equal to 0, construct a data matrix A, construct a matrix E according to the number of categories k, execute the LDA/QR algorithm to obtain a new transfer matrix G, and set the flag to 1; otherwise, execute the ILDA/QR algorithm to update the transfer matrix G , and update the data matrix A;

Step S26: Calculate a new class center vector set C according to the transition matrix G and the data matrix A, form a new classifier, and empty the container;

Step S27: the man-machine training is over; otherwise, return to step S22.

In this embodiment, the mutual trust optimization method described in step S24 includes the following steps:

Step S41: Filter trials according to the time it takes to hit the target; use the time-consuming corresponding to all trials in the container to form a preliminary set, find the minimum time-consuming, and other time-consuming times are generated at their symmetrical positions with the minimum time-consuming time as the symmetric axis A virtual time, all the virtual time is added to the preliminary set to form a reference set, the standard deviation of the reference set is calculated, and the trial that takes less time than the sum of the minimum time spent and the standard deviation is selected;

Step S42: Use the support vector method to screen samples; all samples corresponding to the trial selected in step S41 are added to the candidate sample set, train on the candidate sample set with SVM, and select those samples that are determined to be support vectors;

Wherein, the trial represents an experimental unit in the human-computer training process of the motor imagery brain-computer interface; when the trial starts, the brain-computer interface system randomly sets a target; The system determines the direction to actually move the controlled object by identifying the type of motor imagery; the trial ends when the controlled object hits the target or times out.

In this embodiment, the method for constructing or updating the data matrix A described in steps S12 and S25 is:

Data matrix A=[x ₁ ,x ₂ ,…,x _n ]=[A ₁ ,…,A _k ]∈R ^m×n , x _i ∈R ^m×1 i=1,…,n represents an m-dimensional There are a total of n sample points in A; m is determined by the feature extraction method and is fixed during system operation; n is dynamically increased during system operation; each sample block matrix Represents the set of all sample points of the i-th class, and there are k classes in total, n _i i=1,...,k represents the number of samples of the i-th class,

In this embodiment, the specific method of constructing the matrix E according to the number k of categories described in steps S12 and S25 is as follows:

matrix in $e_i={\left[\begin{array}{cccc}1& 1& ...& 1\end{array}\right]}^T\∈R^{{\mathrm{no}}_i},i=1,...,k.$

In this embodiment, the input of the LDA/QR algorithm is a data matrix A∈R ^m×n and E∈R ^n×k , and the output is a transfer matrix G∈R ^m×k ;

Specifically include the following steps:

Step S31: Calculate the economic QR decomposition of A, A=QR, Q∈R ^m×n , R∈R ^n×n , where Q is column-orthogonal and R is non-singular;

Step S32: Solve a lower triangular linear system R ^T H = E, and obtain H;

Step S33: Calculate G, G=QH.

In this embodiment, the input of the ILDA/QR algorithm is the column orthogonal matrix Q, the optimal transfer matrix G and the new sample x and its category l, and the output is the updated column orthogonal matrix Q and transfer matrix G;

Specifically include the following steps:

Step S41: Calculate r=Q ^T x, Q=[Q(x-Qr)/α];

Step S42: Calculate r=-G ^T x , r(l)=r(l)+1, r=r/α, G=G+Q(:,n+1)r ^T .

In this embodiment, the calculation method of the classification center vector set C is as follows: the data matrix A and the transition matrix G of the existing known categories are calculated P= ^AT G, $P=\left[\begin{array}{c}{p}_1\\ p_2\\ .\\ .\\ .\\ p_{\mathrm{no}}\end{array}\right]=\left[\begin{array}{c}{P}_1\\ P_2\\ .\\ .\\ .\\ P_k\end{array}\right]\∈R^{\mathrm{no}\×k},$ Where p _j ∈ R ^1×k j=1,...,n represents the projection result corresponding to each sample point in the data matrix A, P _i ∈ R ^1×k i=1,...,k represents the i-th class The set of all p _j of ; calculate the center C _i ∈ R ^1×k of each class, $C_i=\frac{1}{{\mathrm{no}}_i}\underset{p_j\∈P_i}{\Σ}{p}_j,$ Get the class center vector set as $C=\left[\begin{array}{c}{C}_1\\ .\\ .\\ .\\ C_k\end{array}\right].$

In this embodiment, the specific classification method of the classifier is: the sample to be determined is x, calculate d _i =||x ^T GC _i || ₂ i=1,...,k, select the smallest d _i , Determine x as the corresponding category.

In this embodiment, the user's motor imagery experiment is composed of multiple runs, and each run is composed of multiple trials, and each trial is composed of multiple overlapping Windowslengths of fixed length. As shown in Figure 4, a trial represents an experimental unit in the human-computer training process of motor imagery BCI, which starts from the appearance of any given target until the controlled object hits the target or times out. Specifically: firstly, at 0s, a target board appears on the screen representing the target that the user is expected to hit, and at this time the user can start to imagine the corresponding movement. After 2s, the controlled object represented by the ball appears in the center of the screen, and in the following 2s-10s, the ball moves accordingly according to the user's imagination determined by the classifier. If the ball hits the target board earlier than 10s, the trial ends early and enters the rest phase. At 10s, no matter whether the ball hits the target board or not, the ball stops moving, the trial ends and enters the rest stage. Both the ball and the target board disappear during the rest phase, and the duration is 2s, after which a new trial starts again. In this embodiment, as shown in Figure 4, the data segment of 2s is used as the length of a fixed-length Windowslength, and the data of Windowslength length is taken every 0.1s, and the Windowslength collected each time is input as the online signal data segment x into the system.

The above descriptions are only preferred embodiments of the present invention, and all equivalent changes and modifications made according to the scope of the patent application of the present invention shall fall within the scope of the present invention.

Claims (10)

1. the two benches adaptive training method of a Mental imagery brain-computer interface, it is characterised in that comprise the following steps:

Step S1: this step is in single trusted phase, Mental imagery brain machine interface system trust data, updates grader by data; Time initial, user carries out the Mental imagery of multiple trial, system online acquisition sample, adopts LDA/QR algorithm to obtain a preliminary transfer matrix G, and calculates classification center vector set C, forms preliminary classification device; User carries on Mental imagery, and system, with the Mental imagery of grader ONLINE RECOGNITION user and to user feedback, obtains new samples simultaneously; Whenever completing a trial, system adopts ILDA/QR algorithm to update transfer matrix G, and calculates classification center vector set C, forms new grader and it is used for identification and the feedback of next trial, until this stage terminates;

Step S2: this step is in mutual trusted phase, Mental imagery brain machine interface system is trust data and grader simultaneously, allows both be mutually adapted; Time initial, system adopts the step S1 Mental imagery of grader ONLINE RECOGNITION user finally given and to user feedback; After user carries out a number of trial, system adopts mutual trust to appoint optimal seeking method screening new samples, LDA/QR algorithm or ILDA/QR algorithm is adopted to update transfer matrix G, and calculate classification center vector set C, form new grader and it is used for ensuing identification and feedback, until man-machine training terminates.

2. the two benches adaptive training method of a kind of Mental imagery brain-computer interface according to claim 1, it is characterised in that described step S1 specifically includes following steps:

Step S11: user carries out the Mental imagery of multiple trial, system online acquisition signal, intercepts a segment signal every an interval, converts m dimensional feature vector to through feature extraction, be designated as x;

Step S12: construct data matrix A by all characteristic vectors of gained in described step S11, and according to classification number k structural matrix E;

Step S13: perform LDA/QR algorithm, obtains the transfer matrix G of optimum, and calculating obtains apoplexy due to endogenous wind Heart vector collection C, forms preliminary classification device;

Step S14: user carries out the Mental imagery of a trial; System online acquisition EEG signals, intercepts a segment signal every an interval and is converted to characteristic vector x through feature extraction, with grader identification x, moves controll plant to user feedback according to recognition result;

Step S15: perform ILDA/QR algorithm, updates transfer matrix G, then updates data matrix A, and calculates class center vector collection C, forms new grader;

Step S16: this stage terminates; Otherwise, new grader is used for identification and the feedback of next trial, returns step S14.

3. the two benches adaptive training method of a kind of Mental imagery brain-computer interface according to claim 1, it is characterised in that described step S2 specifically includes following steps:

Step S21: arranging mark flag is 0, structure container is used for depositing k type games imagination sample;

Step S22: user carries out the Mental imagery of a trial; System online acquisition EEG signals, intercepts a segment signal every an interval and is converted to characteristic vector x through feature extraction, with grader identification x, moves controll plant to user feedback according to recognition result;

All samples of this trial are stored in container by step S23: when current trial terminates, if controll plant hits the mark, and record this trial and hit the mark the spent time;

Step S24: if trial number is not up to adaptive updates condition, then return step S22; Otherwise, system adopts mutual trust to appoint optimal seeking method screening new samples;

Step S25: if flag is equal to 0, construct data matrix A, according to classification number k structural matrix E, performs LDA/QR algorithm and obtains new transfer matrix G, and arranging flag is 1; Otherwise, perform ILDA/QR algorithm and update transfer matrix G, and update data matrix A;

Step S26: calculate according to transfer matrix G and data matrix A and obtain new class center vector collection C, form new grader empty container;

Step S27: man-machine training terminates; Otherwise, step S22 is returned.

4. the two benches adaptive training method of a kind of Mental imagery brain-computer interface according to claim 3, it is characterised in that the mutual trust described in step S24 appoints optimal seeking method to comprise the steps of

Step S41: according to the time screening trial hitting the mark spent; With one preliminary set of composition that expends time in corresponding to trial all in container, find out minimum expending time in, other expend time in and all produce a virtual time into axis of symmetry in its symmetric position with minimum expending time in, all virtual times join preliminary set and form reference set, calculate the standard deviation of reference set, choose and expend time in less than the minimum trial expended time in standard deviation sum;

Step S42: with the method screening sample supporting vector; All sample standard deviations corresponding to the trial that step S41 chooses add candidate samples collection, train on candidate samples collection with SVM, choose those samples being confirmed as supporting vector;

Wherein, described trial represents an experimental considerations unit in the man-machine training process of Mental imagery brain-computer interface; When trial starts, brain machine interface system is a given target at random; User's imagination of launching a campaign tries hard to move controll plant to target, and brain machine interface system is by identifying that the type of Mental imagery determines the direction of actual movement controll plant; When controll plant hits the mark or time-out, trial terminates.

5. the two benches adaptive training method of a kind of Mental imagery brain-computer interface according to claim 2 and 3, it is characterised in that the method for structure described in step S12 and S25 or renewal data matrix A is:

Data matrix A=[x₁,x₂,…,x_n]=[A₁,…,A_k]∈R^m×n, x_i∈R^m×1I × 1 ..., n represents the sample point of a m dimension, total n sample point in A; M is determined by feature extracting method, is changeless in system operation; N dynamically increases; Each sample block matrixWhat represent is the set of all sample points of the i-th class, total k classification, n_iI=1 ..., k represents the number of samples of the i-th class,

6. the two benches adaptive training method of a kind of Mental imagery brain-computer interface according to claim 2 and 3, it is characterised in that described in step S12 and S25 according to classification number k structural matrix E's method particularly includes:

MatrixWherein $e_i={\left[\begin{array}{cccc}1& 1& ...& 1\end{array}\right]}^T\∈R^{n_i},i=1,...,k.$

7. the two benches adaptive training method of a kind of Mental imagery brain-computer interface according to claim 1-3, it is characterised in that the input of described LDA/QR algorithm is data matrix A ∈ R^m×nWith E ∈ R^n×k, it is output as transfer matrix G ∈ R^m×k;

Specifically include following steps:

Step S31: the economic QR calculating A decomposes, A=QR, Q ∈ R^m×n, R ∈ R^n×n, wherein Q is that row are orthogonal, and R is nonsingular;

Step S32: solve a lower triangular linear system R^TH=E, solves and obtains H;

Step S33: calculate and obtain G, G=QH.

8. the two benches adaptive training method of a kind of Mental imagery brain-computer interface according to claim 1-3, it is characterized in that, the input of described ILDA/QR algorithm is row orthogonal matrix Q, optimum transfer matrix G and new sample x and its classification l, is output as the row orthogonal matrix Q and the transfer matrix G that update;

Specifically include following steps:

Step S41: calculate $r=Q^{T}x,\mathrm{\α}=\sqrt{x^{T}x-r^{T}r},$ Q=[Q (x-Qr)/α];

Step S42: calculate r=-G^TX, r (l)=r (l)+1, r=r/ α, G=G+Q (:, n+1) r^T。

9. the two benches adaptive training method of a kind of Mental imagery brain-computer interface according to claim 1-3, it is characterised in that

The computational methods of described classification center vector set C are: the data matrix A and transfer matrix G of existing known class, calculate P=A^TG, $P=\left[\begin{array}{c}{p}_1\\ p_2\\ .\\ .\\ .\\ p_n\end{array}\right]=\left[\begin{array}{c}{P}_1\\ P_2\\ .\\ .\\ .\\ P_k\end{array}\right]\∈R^{n\×k},$ Wherein p_j∈R^1×kJ1 ..., n represents the projection result that in data matrix A, each sample point is corresponding, P_i∈R^1×kI=1 ..., what k represented is belonging to all p of the i-th class_jSet; Calculate the center C of each class_i∈R^1×k,Obtaining class center vector collection is $C=\left[\begin{array}{c}{C}_1\\ .\\ .\\ .\\ C_k\end{array}\right].$

10. the two benches adaptive training method of a kind of Mental imagery brain-computer interface according to claim 1-3, it is characterised in that the concrete sorting technique of described grader is: sample to be determined is x, calculates d_i=| | x^TG-C_i||₂I=1 ..., k, selects minimum d_i, x is judged to the classification of correspondence.