CN106022256A - Parameter optimization method for decision-making model of brain-computer interface system - Google Patents

Abstract

The invention relates to the technical field of brain-computer interfaces, especially a parameter optimization method for a decision-making model of a brain-computer interface system. The method comprises the following steps: (1), collecting brain-computer training data, and carrying out the preprocessing; (2), carrying out data classification through a linear space integrated single detection method, and the positioning of an evidence-accumulating start process. The method employs the linear space integrated single detection method to carry out the classification and recognition of two experiment conditions in training data, carries out the time locating of the evidence-accumulating process in a decision-making model through recognizing the change tendency of the accuracy with time in a single experiment, and carries out the parameter optimization of a sequential decision model. Compared with a previous sequential decision model, the method carries out the locating of the evidence-accumulating process in the decision-making model, eliminates the ineffective classified information accumulation process, carries out the accumulation of the effective classification information in the evidence-accumulating process, and improves the instantaneity of the brain-computer interface system based on the decision-making model.

Description

A kind of parameter optimization method of brain machine interface system decision model

Technical field

The present invention relates to the parameter optimization method of a kind of brain machine interface system decision model, belong to brain-computer interface technology neck Territory.

Background technology

Brain machine interface system is a kind of communication system controlled without relying on muscle, it is intended to damage for disabled patient or motion Hinder patient and a kind of new communications conduit is provided.The core of brain-computer interface technology just by processing the classification of EEG signals, identifies Different conscious activity states.The faintest from the EEG signals of scalp record, signal to noise ratio is the lowest, the most different pattern recognitions Method is applied to brain machine interface system to extract brain electrical feature information, and trains grader to reduce classification error to greatest extent Rate.To this end, brain electricity sorting technique is conducted extensive research by research worker, such as support vector machine, artificial neural network etc..This A little methods achieve preferable classifying quality in its field being suitable for, but only brain electrical feature to special time period is carried out point Class, does not accounts for the temporal information in categorizing process.EEG signals has non-stationary, the differentiation of each time period brain electrical feature Degree is the most different, and therefore these graders cannot weigh the relation between classification accuracy and decision-making time well.For reality Time the training set Limited Number that can gather of brain machine interface system, how to go out the grader of better performances with a small amount of sample training, i.e. " small sample problem " is the challenge that brain machine interface system faces.Therefore, attention is gradually gone to by Recent study personnel The dynamic cataloging method of user view can be reflected continuously.

" classification " in brain-computer interface field generally means that decision-making or selecting response.Decision-making is the one of biological nervous system Higher cognitive process, its Information procession be unique in that can on transient state stimulate rapidly make judgement thus affect behavior.By Information in decision making process sensation often has uncertainty, and therefore statistical theory is the strong of formalized description uncertain information Broad theory instrument.Sequential analysis theory provides a mathematical model flexibly to decision making process.Sequential analysis technique study is certainly During question and answer on politics topic, not being fixed sample amount in advance, but gradually sample, its required average sample size is minimum, advantageously accounts for " small sample problem " of brain machine interface system, has the most gradually attracted the attention of brain machine interface system research worker.

In sequential decision model, decision making process is exactly the letter that noisy information starts to arrive accumulation border from accumulation starting point Breath cumulative process, when information accumulation reaches certain border, then it is assumed that decision-making completes.Therefore, sequential likelihood ratio test sorting technique Utilize probability ratio test method to calculate the classification information of every section of feature, start iterated integral category information from data starting point, gradually take Sample, until judging when accumulation decision variable arrives a certain threshold value, stops accumulation.The method can reflect that decision making process is to propping up Hold the balance between adaptation response and speed of decision and the accuracy selecting evidence to make, be therefore applicable to online brain machine and connect Port system.But, nowadays sequential likelihood ratio test sorting technique is all just to carry out the tired of brain electrical feature information when experiment starts Long-pending, a length of whole segment data length of its maximum accumulation time window.The research of consciousness decision domain both at home and abroad shows, the process bag of decision-making Include stimulation perception, it is noted that the processes such as evidence accumulation and Motor execution.Stimulating perception, noticing etc. that the stage carries out decision information Accumulation be unhelpful to classification results, add the decision-making time of brain machine interface system on the contrary, reduce the real-time of system. Therefore, evidence cumulative process is positioned, specify the starting point of effective information accumulation, it is possible to reduce based on sequential decision model The amount of calculation of dynamic classifier, the maximum accumulation time window shortening algorithm is long, to improve the real-time performance of brain machine interface system.

Summary of the invention

In order to overcome problems of the prior art, it is an object of the present invention to provide a kind of brain machine interface system decision model Parameter optimization method.The present invention utilizes linear space integration single probe technique to accumulate the evidence during decision-making and initiateed Cheng Jinhang positions, thus optimizes brain machine interface system decision model parameter further.The method was by initiateing evidence accumulation Cheng Jinhang positions, and gets rid of dynamic classifier and carries out the time loss of brain electrical feature accumulation at unrelated procedures, not only reduces sequential The accumulated time of information, and improve the efficiency of decision-making, thus improve the real-time of brain machine interface system further.

In order to realize foregoing invention purpose, solve the problem in the presence of prior art, the technical scheme that the present invention takes It is: the parameter optimization method of a kind of brain machine interface system decision model, comprises the following steps:

Step 1, collection EEG signals training data also carry out pretreatment, utilize brain wave acquisition equipment to gather random point motion EEG signals under direction discernment task is as training data, and different according to the direction of motion and experiment difficulty, this random point is transported Dynamic direction includes low difficulty difficulty left, low difficulty right, middle difficulty left, a middle left side right, highly difficult and highly difficult right six kinds of different identifications Task, and the EEG signals collected is filtered, removes artefact pretreatment；According to random point direction of motion label to collection EEG signals carry out segment processing, and in order to distinguish different experiment examinations time, split time difficulty same for same experimenter The different experiments examination time of degree needs to keep consistent, and the split time length taken needs to comprise to stimulate and occurs to making a response Process, i.e. experimenter carry out the process of decision-making；

The classification of step 2, linear space integration single probe technique data and evidence accumulation initiating process location, by linearly Two kinds of data categories in experiment are identified, the such as identification to the direction of motion, Zuo Huo by space integration single probe technique The right side, the identification to experiment difficulty, difficult or easily, specifically include following sub-step:

A () structural classification initial data X, classification initial data X should include two class data, the most left and right direction of motion two Class data, i.e. X₁And X₂, wherein, X₁It is a n_channels×T_sample×N₁Data matrix, X₂It is a n_channels×T_sample ×N₂Data matrix, wherein N₁With N₂Represent the experiment examination number of times that two class packets contain, T respectively_sampleFor what step 1 used The sampling number of split time, n_channelsFor the number that leads used, and the initial data X=[X that classifies₁,X₂], it is a n_channels ×T_sample×(N₁+N₂) matrix, and the classification initial data X-form obtained is converted into (T_sample×N_trials)× n_channelsMatrix form in the ensuing Classification and Identification, N_trialsAll by carry out that the data of Classification and Identification are comprised Experiment examination number of times, N_trials=N₁+N₂；

B () structure sliding window, the window width of sliding window is δ sampled point, and single sliding distance is τ sampled point, for adopting Number of samples is T_sampleData can be divided into altogetherIndividual sliding window, whereinThe meaning rounded under expression Thinking, the position of each sliding window is [1+ (i-1) τ, δ+(i-1) τ] respectively, i=1,2 ..., K, in utilizing window, data are to dividing Class device is trained, and is carried out the time starting point of positive evidence accumulation by the variation tendency of classification accuracy；

C () structure class mark sequence L, according in the original brain electricity categorical data obtained in sub-step (a) and sub-step (b) Sliding window, structure comprises element { class mark sequence L of 0,1}, L is (δ × a N_trialsThe column vector of) × 1, wherein δ is for sliding The window of dynamic window is long, N_trialsBy carrying out all experiment examination number of times that the data of Classification and Identification are comprised；

D () uses logistic regression linear classifier, estimate the classification initial data in each sliding window built in step (b) Count an optimal spatial weighted vector w_τ,δ, two class data in this window can be carried out maximizing identification, pass through formula by these weights (1) classification results y it is calculated,

Y=w^TX+b (1)

In formula, w is the space weighted vector of this linear classifier, w^TRepresenting the transposed matrix of w, b is bias term, and X is for dividing Class initial data, it is assumed that probability sorted out by the sample of classification initial data meets formula (2),

$\left\{\begin{array}{c}p(c=+1|X)=f\left(y\right)=\frac{1}{1+e^{-y}}=\frac{1}{1+e^{-(w^{T}X+b)}}\\ p(c=-1|X)=1-f\left(y\right)\end{array}\right.---\left(2\right)$

In formula, p (c=+1 | X) represents that X is judged as the probability of class c=+1, and p (c=-1 | X) represent that X is judged as class The probability of c=-1, obtains the space weight w of optimum by iteration weight weighted least-squares method, and concrete iterative process is by public affairs Formula (3) is realized,

$\begin{array}{c}p\←f(w^T*X)\\ g\←X*(d-p)-\mathrm{\Λ}*w\\ H\←X*diag(p.*(1-p\left)\right)*X^T+\mathrm{\Λ}*I\\ w\←w+H^{-1}*g\end{array}---\left(3\right)$

In formula, X is classification initial data, X^TFor the transposition of X, p is the classification probability vector of sample, and * represents inner product .* generation Table vector product, d represents and comprises brain electricity sample class target column vector, i.e. L in sub-step (c), diag () represent a vector Being converted into diagonal matrix, g represents gradient vector, and H is the gloomy battle array in sea that Fisher information matrix obtains, and Λ is punishing in iterative process Penalty factor, during performing, during bias term b of linear classification is integrated into space weight w originally, is retouched by formula (4) State,

$\begin{array}{c}\[w,b\]\→w\\ \[X,1\]\→X\\ \mathrm{\Λ}\→\mathrm{\λ}\left[\begin{array}{cc}1& 0\\ 0& 0\end{array}\right],\mathrm{\λ}={10}^{-6}\end{array}---\left(4\right)$

Space weight w is initialized as 10^-3*ones(n_channels+ 1,1), wherein ones (n_channels+ 1,1) one is represented (n_channels+ 1) all 1's matrix of × 1, n_channelsRepresenting the number that leads being acquired, iteration process is up to space weights Till w convergence, thus obtain the space weight w of optimal classification, be 1+ (i-owing to whole categorizing process is aimed at starting point 1) τ, i=1,2 ..., K, data in the sliding window of a length of δ of window, the most now optimal classification space weight w is considered as w_τ,δ, whereinExpression sampling number is T_sampleThe sliding window number that separates of data, τ represents slip every time Distance；

The e space weight w of optimal classification that () will ask in sub-step (d)_τ,δBring in formula (1), obtain classification results Y, and utilize formula (2) to calculate the probability distribution of y, utilize the probability distribution result of classification results y to construct with sub-step (c) simultaneously Class mark sequence draw susceptibility curve, i.e. ROC curve, and calculate the area Az under ROC curve；

F classification results Az that data in time windows are obtained by () adds up, and draws the song of Az window in time change Line, due to the beginning accumulated along with evidence, the classifying quality to two generic tasks should be the trend presenting a kind of rising, i.e. evidence Accumulating the most more sufficient, the accuracy made accurate judgment is the highest, and ought to reach maximum when decision-making performs；Therefore, to Az The curve of window change in time is analyzed, and observes the start time of its lasting ascent stage, starts in order to conclusion evidence accumulation Moment, and this moment is started as the information characteristics accumulation of next online brain machine interface system based on sequential decision model Moment, thus get rid of the unrelated classification information of accumulation, only effective classification information is accumulated, thus minimizing processes the time, reaches Purpose to the real-time improving brain machine interface system.

The medicine have the advantages that the parameter optimization method of a kind of brain machine interface system decision model, comprise the following steps: (1) gather EEG signals training data and carry out pretreatment, the classification of (2) linear space integration single probe technique data and evidence Accumulation initiating process positions.Compared with the prior art, the present invention uses linear space integration single detection method, to training number Two kinds of experimental conditions according to carry out Classification and Identification, the accuracy rate identified by analysis change over time during single experiment The trend changed, carries out temporal location to the evidence cumulative process in decision model, thus optimizes the letter of sequential decision model Breath accumulation parameter, compared with previous sequential decision model, this invention is by carrying out the evidence cumulative process in decision model Location, eliminates invalid classifications information accumulation process, shortens maximum accumulation time window long, improve based on this decision model The real-time of online brain machine interface system.

Accompanying drawing explanation

Fig. 1 is the inventive method flow chart of steps.

Fig. 2 is the random point side-to-side movement direction discernment experimental paradigm figure that brain wave acquisition of the present invention is used.

Fig. 3 is the variation tendency that in the present invention, all subject motion's direction discernment accuracy rate averages are slided along with sliding window Figure.

Fig. 4 is the variation tendency that in the present invention, all subject's difficulty recognition accuracy averages are slided along with sliding window Figure.

Detailed description of the invention

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

As it is shown in figure 1, the parameter optimization method of a kind of brain machine interface system decision model, comprise the following steps:

Step 1, gathers 10 experimenter's EEG signals training datas and carries out pretreatment, utilizing brain wave acquisition equipment NeuroScan and 36 conduction polar cap gathers the brain telecommunications under the random point side-to-side movement direction discernment task of three kinds of experiment difficulty Number as training data, brain electricity sample rate is 500Hz, and arranges the band filter of 0.1-70Hz.Wherein brain wave acquisition experiment Normal form, as in figure 2 it is shown, need experimenter to watch screen attentively in gatherer process, is tested after a sound " Beep " is pointed out and is started, prompting sound 1.5s after end, random point stimulates appearance.Random point stimulation includes two parts, and the first is unanimously moved to same direction Random point；It two is the random point to all directions random motion.When random point occurs, experimenter needs to identify as early as possible to same The direction of motion of the random point unanimously moved in one direction, and according to judgement, make the reaction of correspondence, such as right-hand man's button is anti- Should, the most tested to make a response, then random point disappears, and stimulates after disappearing, and subjects carries out the rest of 2 seconds, enters and tries next time Test.The ratio testing the random point by changing consistent motion sets three kinds of different experiment difficulty.Therefore the brain that will gather Electricity data divide into six classes: low difficulty is left, and low difficulty is right, and middle difficulty is left, and middle difficulty is right, a highly difficult left side, the highly difficult right side.It is directed to Each difficulty, is circulated by above flow process, and each difficulty gathers the data of 210 correct responses.After gathering signal, to adopting Collect to EEG signals carry out pretreatment, remove most of high frequencies and flip-flop including the wave filter utilizing 0.1-30Hz, Then using independent component analysis, the method such as threshold method removes the artefacts such as eye electricity.Subsequently, according to random point direction of motion label pair Pretreated EEG signals carries out segment processing, and in order to distinguish different experiment examinations time, split time is for same experimenter The different experiments examination time of same difficulty needs to keep consistent, and the split time length taken needs to comprise stimulation and occurs to making The process of reaction, i.e. experimenter carries out the process of decision-making；After data are carried out segmentation, enter step 2.

Step 2, the classification of linear space integration single probe technique data and evidence accumulation initiating process location, by linearly Two kinds of data categories in experiment are identified, the such as identification to the direction of motion, Zuo Huo by space integration single probe technique The right side, the identification to experiment difficulty, difficult or easily, specifically include following sub-step:

A () structural classification initial data X, classification initial data X should include that two class data, two class data are defined as X₁And X₂, Wherein, X₁It is a n_channels×T_sample×N₁Data matrix, X₂It is a n_channels×T_sample×N₂Data matrix, Wherein N₁With N₂Represent the experiment examination number of times that two class packets contain, T respectively_sampleThe sampling of the split time for using in step 1 Count, n_channelsFor the number that leads used, and the initial data X=[X that classifies₁,X₂], it is a n_channels×T_sample×(N₁+ N₂) matrix, and the classification initial data X-form obtained is converted into (T_sample×N_trials)×n_channelsMatrix form use In ensuing Classification and Identification, N_trialsBy carrying out all experiment examination number of times that the data of Classification and Identification are comprised, N_trials= N₁+N₂；

The present invention comes to two kinds and identifies content, be direction of motion identification and experiment difficulty identification respectively, therefore classify Initial data has two classes:

For direction of motion identification: label based on random point, the eeg data of segmentation is divided into two classes X_L、X_R, respectively Represent left direction data and right direction data, X the most above₁、X₂.It is reconstructed again and obtains new brain electricity categorical data X=[X_L, X_R], it is a n_channels×T_sample×(N_L+N_R) matrix of size.Wherein T_sampleFor the split time of employing in step 1 Sampling number, n_channelsFor the number that leads used, N_LAnd N_RRepresent the experiment number of left direction and right direction respectively.

For difficulty level identification: three kinds of experiment difficulty are divided into two kinds of difficulty identification missions, the lowest difficulty and middle difficulty Identification, low difficulty and highly difficult identification.First by brain electricity number corresponding with middle difficulty for the low difficulty in three kinds of experiment difficulty According to taking-up, combination obtains new eeg data X=[X_LD,X_MD], wherein X_LD、X_MDRepresent low difficulty data and middle difficulty number respectively According to, X the most above₁、X₂.This stylish eeg data X is a n_channels×T_sample×(N_LD+N_MD) matrix of size.Wherein T_sampleThe sampling number of the split time for using in step 1, n_channelsFor the number that leads used, N_LD、N_MDRepresent low respectively Difficulty and the experiment number of middle difficulty.

Secondly the low difficulty in three kinds of experiment difficulty being taken out with highly difficult corresponding eeg data, combination obtains another New eeg data X=[X_LD,X_HD], wherein X_LD、X_HDRepresent low difficulty data and highly difficult data, the most above X respectively₁、X₂。 This stylish eeg data X is a n_channels×T_sample×(N_LD+N_HD) matrix of size.Wherein T_sampleFor step 1 is adopted The sampling number of split time, n_channelsFor the number that leads used, N_LD、N_HDRepresent low difficulty and highly difficult reality respectively Test number of times.

(b) structure sliding window；The window width of sliding window is δ sampled point, and single sliding distance is τ sampled point, for adopting Number of samples is T_sampleData can be divided into altogetherIndividual sliding window, whereinThe meaning rounded under expression Think.The position of sliding window is [1+ (i-1) τ, δ+(i-1) τ] respectively every time, i=1,2 ..., K.In utilizing window, data are to dividing Class device is trained, and is carried out the time starting point of positive evidence accumulation by the variation tendency of classification accuracy.Sliding window of the present invention Window width is set as δ=30 sampled point, single sliding distance τ=15 sampled point.

C (), constructs class mark sequence L；According to the original brain electricity categorical data obtained in sub-step (a) and sub-step (b) In sliding window, structure comprises element { class mark sequence L of 0,1}, L is (δ × a N_trialsThe column vector of) × 1.Now δ is The window of sliding window is long, N_trialsBy carrying out all experiment examination number of times that the data of Classification and Identification are comprised.

D () logistic regression estimates optimal spatial weighted vector；Use logistic regression, to each slip built in step (b) Classification initial data in window estimates an optimal spatial weighted vector w_τ,δ, two class data in this window can be carried out by these weights Maximizing and identify, wherein logistic regression is a kind of linear classifier, is calculated classification results y by formula (1),

Y=w^TX+b (1)

In formula, w is the space weighted vector of this linear classifier, w^TRepresenting the transposed matrix of w, b is bias term, and X is for dividing Class initial data, it is assumed that probability sorted out by the sample of classification initial data meets formula (2),

$\left\{\begin{array}{c}p(c=+1|X)=f\left(y\right)=\frac{1}{1+e^{-y}}=\frac{1}{1+e^{-(w^{T}X+b)}}\\ p(c=-1|X)=1-f\left(y\right)\end{array}\right.---\left(2\right)$

In formula, p (c=+1 | X) represents that X is judged as the probability of class c=+1, and p (c=-1 | X) represent that X is judged as class The probability of c=-1, obtains the space weight w of optimum by iteration weight weighted least-squares method, and concrete iterative process is by public affairs Formula (3) is realized,

$\begin{array}{c}p\←f(w^T*X)\\ g\←X*(d-p)-\mathrm{\Λ}*w\\ H\←X*diag(p.*(1-p\left)\right)*X^T+\mathrm{\Λ}*I\\ w\←w+H^{-1}*g\end{array}---\left(3\right)$

In formula, X is classification initial data, X^TFor the transposition of X, p is the classification probability of sample, and * represents inner product, and .* represents arrow Amount is long-pending, and d represents that one comprises each brain electricity sample class mark { L, diag () in the column vector of 1,0}, i.e. step (3) can be by one Individual vector median filters diagonally matrix, column vector g represents gradient, and H is the gloomy battle array in sea that Fisher information matrix obtains, and Λ is iteration mistake Penalty factor in journey.During performing, during bias term b of linear classification is integrated into space weight w originally, therefore exist Above-mentioned iterative process does not occur, is described by formula (4),

$\begin{array}{c}\[w,b\]\→w\\ \[X,1\]\→X\\ \mathrm{\Λ}\→\mathrm{\λ}\left[\begin{array}{cc}1& 0\\ 0& 0\end{array}\right],\mathrm{\λ}={10}^{-6}\end{array}---\left(4\right)$

Space weight w is initialized as 10^-3*ones(n_channels+ 1,1), wherein ones (n_channels+ 1,1) one is represented (n_channels+ 1) all 1's matrix of × 1, n_channelsRepresenting the number that leads being acquired, iteration process is up to space weights Till w convergence, thus obtain the space weight w of optimal classification, be 1+ (i-owing to whole categorizing process is aimed at starting point 1) τ, i=1,2 ..., K, data in the sliding window of a length of δ of window, the most now optimal classification space weight w is considered as w_τ,δ, whereinExpression sampling number is T_sampleThe sliding window number that separates of data, τ represents slip every time Distance；

E () draws susceptibility curve, i.e. ROC curve solve area under curve according to classification results；By in sub-step (d) The space weight w of the optimal classification asked for_τ,δSubstitute into y=w^TIn X+b, obtain classification results y, and utilize formula (2) to calculate y's Probability distribution, utilizes the probability distribution result of classification results y to draw ROC curve with the class mark sequence that sub-step (c) constructs simultaneously, And calculate the area Az under ROC curve；

F () Az analysis of trend also positions evidence cumulative process；The classification results that data in time windows are obtained Az adds up, and draws the curve of Az window in time change.Wherein the curve of direction of motion identification is as it is shown on figure 3, and test difficulty Degree identifies curve as shown in Figure 4.Due to the beginning accumulated along with evidence, the classifying quality to two generic tasks should be to present one Rise trend, i.e. evidence accumulation the most more sufficient, the accuracy made accurate judgment is the highest, and ought to when decision-making performs, Reach maximum；Therefore, the curve of Az window in time change is analyzed, finds that the time observing its lasting ascent stage rises Point, accumulates start time in order to conclusion evidence, and is connect as next online brain machine based on sequential decision model in this moment The information characteristics accumulation start time of port system.Can be seen that two articles of curves, Az is all from the 7th sliding window from Fig. 3 and Fig. 4 Position start to present the variation tendency persistently risen, corresponding to time shaft greatly occur approximately in stimulation occur after about 210ms, because of This will stimulate after 210ms as the starting point of maximum accumulation time window, thus unrelated classification information before getting rid of this moment point, the most right Effectively classification information is accumulated, and thus minimizing processes the time, reaches to improve the purpose of the real-time of brain machine interface system.

Claims (1)

1. the parameter optimization method of a brain machine interface system decision model, it is characterised in that comprise the following steps:

Step 1, collection EEG signals training data also carry out pretreatment, utilize brain wave acquisition equipment to gather the random point direction of motion EEG signals under identification mission is as training data, different according to the direction of motion and experiment difficulty, this random point motion side To including low difficulty difficulty left, low difficulty right, middle difficulty left, a middle left side right, highly difficult and highly difficult right six kinds of different identification missions, And the EEG signals collected is filtered, removes artefact pretreatment；According to the random point direction of motion label brain to gathering The signal of telecommunication carries out segment processing, and in order to distinguish different experiment examinations time, split time difficulty same for same experimenter Different experiments examination time needs to keep consistent, and the split time length taken needs to comprise stimulation and occurs to the mistake made a response Journey, i.e. experimenter carry out the process of decision-making；

The classification of step 2, linear space integration single probe technique data and evidence accumulation initiating process location, pass through linear space Two kinds of data categories in experiment are identified, the such as identification to the direction of motion, left or right by integration single probe technique, right The identification of experiment difficulty, difficult or easily, specifically include following sub-step:

A () structural classification initial data X, classification initial data X should include two class data, the most left and right direction of motion two class number According to, i.e. X₁And X₂, wherein, X₁It is a n_channels×T_sample×N₁Data matrix, X₂It is a n_channels×T_sample×N₂ Data matrix, wherein N₁With N₂Represent the experiment examination number of times that two class packets contain, T respectively_sampleFor the segmentation used in step 1 The sampling number of time, n_channelsFor the number that leads used, and the initial data X=[X that classifies₁,X₂], it is a n_channels× T_sample×(N₁+N₂) matrix, and the classification initial data X-form obtained is converted into (T_sample×N_trials)×n_channels Matrix form in the ensuing Classification and Identification, N_trialsBy carrying out all experiment examinations that the data of Classification and Identification are comprised Number of times, N_trials=N₁+N₂；

B () structure sliding window, the window width of sliding window is δ sampled point, and single sliding distance is τ sampled point, for sampled point Number is T_sampleData can be divided into altogetherIndividual sliding window, whereinThe meaning rounded under expression, often The position of secondary sliding window is [1+ (i-1) τ, δ+(i-1) τ] respectively, i=1,2 ..., K, in utilizing window, data are to grader It is trained, is carried out the time starting point of positive evidence accumulation by the variation tendency of classification accuracy；

C () structure class mark sequence L, according to the cunning in the original brain electricity categorical data obtained in sub-step (a) and sub-step (b) Dynamic window, structure comprises element, and { class mark sequence L of 0,1}, L is (δ × a N_trialsThe column vector of) × 1, wherein δ is sliding window Window long, N_trialsBy carrying out all experiment examination number of times that the data of Classification and Identification are comprised；

D () uses logistic regression linear classifier, the classification initial data in each sliding window built in step (b) is estimated one Individual optimal spatial weighted vector w_τ,δ, two class data in this window can be carried out maximizing identification, be counted by formula (1) by these weights Calculation obtains classification results y,

Y=w^TX+b (1)

In formula, w is the space weighted vector of this linear classifier, w^TRepresenting the transposed matrix of w, b is bias term, and X is original for classification Data, it is assumed that probability sorted out by the sample of classification initial data meets formula (2),

$\left\{\begin{array}{c}p(c=+1|X)=f\left(y\right)=\frac{1}{1+e^{-y}}=\frac{1}{1+e^{-(w^{T}X+b)}}\\ p(c=-1|X)=1-f\left(y\right)\end{array}\right.---\left(2\right)$

In formula, p (c=+1 | X) represents that X is judged as the probability of class c=+1, and p (c=-1 | X) represent that X is judged as class c=- The probability of 1, obtains the space weight w of optimum by iteration weight weighted least-squares method, and concrete iterative process passes through formula (3) Realized,

$\begin{array}{c}p\←f(w^T*X)\\ g\←X*(d-p)-\mathrm{\Λ}*w\\ H\←X*diag(p.*(1-p\left)\right)*X^T+\mathrm{\Λ}*I\\ w\←w+H^{-1}*g\end{array}---\left(3\right)$

In formula, X is classification initial data, X^TFor the transposition of X, p is the classification probability vector of sample, and * represents inner product, and .* represents arrow Amount is long-pending, and d represents and comprises brain electricity sample class target column vector, i.e. L in sub-step (c), diag () represent a vector median filters Diagonally matrix, g represents gradient vector, and H is the gloomy battle array in sea that Fisher information matrix obtains, Λ be punishment in iterative process because of Son, during performing, during bias term b of linear classification is integrated into space weight w originally, is described by formula (4),

$\begin{array}{c}\[w,b\]\→w\\ \[X,1\]\→X\\ \mathrm{\Λ}\→\mathrm{\λ}\left[\begin{array}{cc}1& 0\\ 0& 0\end{array}\right],\mathrm{\λ}={10}^{-6}\end{array}---\left(4\right)$

Space weight w is initialized as 10^-3*ones(n_channels+ 1,1), wherein ones (n_channels+ 1,1) one is represented (n_channels+ 1) all 1's matrix of × 1, n_channelsRepresenting the number that leads being acquired, iteration process is up to space weights Till w convergence, thus obtain the space weight w of optimal classification, be 1+ (i-owing to whole categorizing process is aimed at starting point 1) τ, i=1,2 ..., K, data in the sliding window of a length of δ of window, the most now optimal classification space weight w is considered as w_τ,δ, whereinExpression sampling number is T_sampleThe sliding window number that separates of data, τ represents slip every time Distance；

The e space weight w of optimal classification that () will ask in sub-step (d)_τ,δBring in formula (1), obtain classification results y, and Utilize formula (2) to calculate the probability distribution of y, utilize the class that the probability distribution result of classification results y constructs with sub-step (c) simultaneously Mark sequence draws susceptibility curve, i.e. ROC curve, and calculates the area Az under ROC curve；

F classification results Az that data in time windows are obtained by () adds up, and draws the curve of Az window in time change, by In along with evidence accumulation, the classifying quality to two generic tasks should be the trend presenting a kind of rising, i.e. evidence accumulation The most more sufficient, the accuracy made accurate judgment is the highest, and ought to reach maximum when decision-making performs；Therefore, to Az at any time Between window change curve be analyzed, observe the start time of its lasting ascent stage, in order to conclusion evidence accumulate start time, And this moment is accumulated start time as the information characteristics of next online brain machine interface system based on sequential decision model, Thus get rid of the unrelated classification information of accumulation, and only effective classification information is accumulated, thus minimizing processes the time, reaches to improve The purpose of the real-time of brain machine interface system.