CN106108894A - A kind of emotion electroencephalogramrecognition recognition method improving Emotion identification model time robustness - Google Patents

Abstract

The invention discloses a kind of emotion electroencephalogramrecognition recognition method improving Emotion identification model time robustness, including: 64 collected are led EEG signals, and to include that change is referenced to ears average；It is downsampled to 500Hz；1 100Hz bandpass filtering；And utilize the algorithm of independent component analysis to remove the pretreatment of eye electrical interference；By pretreated EEG signals by optimal separability frequency range can be found by frequency-division section adaptive tracing, calculate respectively each lead most preferably can the power spectral density of frequency-division section, constitute emotional characteristics matrix；Utilize PCA that eigenmatrix is carried out dimensionality reduction；Use support vector machine classifier that the electroencephalogram power spectrum signature after dimensionality reduction is identified, set up Emotion identification model.The present invention, by finding optimal separability frequency range by frequency-division section adaptive tracing, enhances emotion correlated characteristic by the natural law of sample in the training set of increase Emotion identification model, weakens temporal feature, improve the time robustness of Emotion identification model.

Description

A kind of emotion electroencephalogramrecognition recognition method improving Emotion identification model time robustness

Technical field

The present invention relates to electroencephalogramrecognition recognition field, particularly relate to a kind of emotion brain improving Emotion identification model time robustness Electricity recognition methods.

Background technology

Emotion (emotion) is whether people meets self needs to objective things and the comprehensive state that produces.It is as people The Premium Features of brain, it is ensured that organic existence and adaptation, affect the study of people, memory and decision-making in varying degrees.? In the routine work of people and life, the effect of emotion is ubiquitous.Negative Emotional can affect our physical and mental health, reduces work Make quality and efficiency, will also result in serious work mistake.There are some researches prove that the long term accumulation of Negative Emotional can damage immunity The function of system, makes people be infected more susceptible to virus around.So, in time find Negative Emotional and give suitable Intervene the most necessary, especially to some particular job persons such as driver, spacefarer with regulation and control.On the other hand, in man-machine interaction system In system, if system can capture the emotional state of people, then man-machine interaction will become more friendly, natural and efficient. The analysis of emotion and identification have become as the fields such as neuroscience, psychology, Cognitive Science, computer science and artificial intelligence and learn The important research topic that section intersects.

Development and the rise of brain imaging technique, EEG signals along with neuro physiology (Electroencephalography, EEG) because its temporal resolution is high, not by anthropic factor control, can objective reality ground The emotional state of reflection people and by the attention of research worker and be introduced in Emotion identification field.And new-type theoretical method Propose to improve Emotion identification accuracy rate based on brain electricity to a certain extent.But once moving towards actual application, discrimination is big Amplitude declines, and is difficult to meet the demand of application, and the Emotion identification model setting up pinpoint accuracy still faces huge challenge.

One of them difficult point is how to reject or reduce the time effect of EEG signals, and then improves Emotion identification model Time universality.It is known that hormonal readiness, external environment condition (such as temperature and humidity), and diet can draw with sleep Rise physiological signal difference, even if so in the different time EEG signals under same emotional state be also variant Property.And in actual applications, will certainly there is the regular hour in the foundation of Emotion identification model and the identification of emotional state Interval, and test data and cannot participate in the foundation of Emotion identification model, particularly in the application scenarios that some are special, such as The identification of spacefarer's emotional state, identify model sets up the preparatory stage occurred on the ground, and the identification of emotional state is sent out Raw working stage in space.It is unpractical for setting up the same day and identifying that then model immediately enters application.

In sum, reject or reduce the time effect of EEG signals, and then improve the time robust of Emotion identification model Property is the most necessary.In existing research, the research about the time universality of emotion grader is very few.2001 Year, Picard et al.^[1]Attempt removing the time effect impact on Emotion identification model, use other emotional states to deduct calmness The method of state, but utilize the method can reduce with regard to None-identified neutrality emotion, type of emotion, and centering emotional state Identifying it is also extremely important and indispensable, neutral emotion is an important indicator of emotional stability.2012, Chueh, Tung-Hung et al.^[2]The method utilizing multivariate analysis of variance removes the impact of time effect, improves the performance of grader.But It is still to there is a problem, it is simply that the data in test set are not independent, and still the data mixing with other times exists Building grader together, this is the most also unpractical.

Summary of the invention

The invention provides a kind of emotion electroencephalogramrecognition recognition method improving Emotion identification model time robustness, the present invention can It is effectively improved time robustness and the universality of Emotion identification model, solves the bottleneck problem in current Emotion identification, by mould Application pushed to by type, and obtains considerable Social benefit and economic benefit, described below:

A kind of emotion electroencephalogramrecognition recognition method improving Emotion identification model time robustness, described emotion electroencephalogramrecognition recognition method Comprise the following steps:

EEG signals is led in 64 collected and carries out pretreatment, including: it is average that change is referenced to ears；It is downsampled to 500Hz； 1-100Hz bandpass filtering；And utilize the algorithm of independent component analysis to remove eye electrical interference；

Pretreated EEG signals is used and the algorithm of frequency-division section adaptive tracing can find most preferably may be used of each user Point property frequency range, calculate respectively each lead most preferably can the power spectral density of frequency-division section, constitute emotional characteristics matrix；

PCA is utilized the emotional characteristics matrix obtained to be carried out dimension-reduction treatment, as final eigenmatrix；

Use support vector machine classifier that the feature in final eigenmatrix is identified, by increasing mood model In training set, the natural law of sample weakens temporal characteristics, improves the time robustness of mood model, is distinguished by difference emotional state, Set up Emotion identification model.

Described method also includes:

Gather different time sections subjects 64 EEG signals led under different emotional states.

Described pretreated EEG signals employing the algorithm of frequency-division section adaptive tracing can find each user The step of good separability frequency range particularly as follows:

1) Short Time Fourier Transform is used to calculate each time-frequency matrix led；

2) fisher ratio, the capacity volume variance in weighing class with it and between class are calculated；

3) can fraction weight DW (f) can be tried to achieve by Fisher ratio；DFC, iterations is calculated by wave band iteration selection method Equal to the band number needing acquisition；Optimal separability frequency range is obtained from the band number obtained.

Described from the band number obtained, obtain optimal separability frequency range particularly as follows:

When frequency window moves along the frequency axis of DW (f), calculate Energy distribution；It is distributed α, in institute according to ceiling capacity Some band numbers select optimalCalculateRelatively change δ_j；

If a threshold value δ_min, compare δ₂And δ_minIf, δ₂More than δ_min, the most then compare δ₃With δ_minSize, until Find δ_jLess than δ_min, then the position of j-1 is exactly the frequency range of optimal separability.

Described utilize PCA that the emotional characteristics matrix obtained is carried out dimension-reduction treatment, as final feature square Battle array step particularly as follows:

1) it is standardized initial data processing, obtains original matrix；Then its covariance matrix is sought；To covariance square Battle array carries out characteristic root decomposition, obtains characteristic root matrix and characteristic vector；

2) original matrix projection in new vector space, i.e. principal component vector group are asked；

3) the characteristic root size of each main constituent represents its number containing quantity of information, asks the accumulation tribute of front k main constituent Offer rate；

4) the selected accumulation contribution rate preset, makes front d main constituent F_Ni*dPattern recognition is carried out as new data.

Feature in final eigenmatrix is identified by described use support vector machine classifier, by difference emotion shape State distinguishes, set up the step of Emotion identification model particularly as follows:

The data of every day are carried out row normalization respectively, normalizes to [-1.1], obtain eigenmatrix；

Utilize SVM classifier to set up Emotion identification model, in modeling process, the data of many days are put into training set, to carry The time robustness of high-class device.

The technical scheme that the present invention provides provides the benefit that: the purport of the present invention is to propose one to improve Emotion identification mould The new method of type time robustness, by can frequency-division section adaptive tracing method search out each user most preferably can frequency-division section, logical Crossing and improve in training set that the natural law of sample is to weaken temporal feature, accurate stable carries out Emotion identification in real time then. This invention can be effectively improved time robustness and the accuracy of Emotion identification model, can obtain considerable social benefit and warp Ji benefit.Optimum implementation is intended using patent transfer, technological cooperation or product development.

Accompanying drawing explanation

Fig. 1 is the flow chart of a kind of emotion electroencephalogramrecognition recognition method improving Emotion identification model time robustness；

Fig. 2 is 60 to lead EEG and lead figure；

Fig. 3 is tested experimental period calendar；

Fig. 4 is can frequency-division section adaptive tracing calculation flow chart；

Fig. 5 is DFCs algorithm flow chart；

Fig. 6 is the recognition correct rate under different training natural law.

Detailed description of the invention

For making the object, technical solutions and advantages of the present invention clearer, below embodiment of the present invention is made further Ground describes in detail.

In order to solve the problem in background technology, when the embodiment of the present invention proposes a kind of new raising Emotion identification model Between the emotion electroencephalogramrecognition recognition method of robustness, by the optimal separability of each user can be found by frequency-division section adaptive tracking algorithm Frequency range, calculate respectively each lead most preferably can the power spectral density of frequency-division section, constitute emotional characteristics matrix, utilize main constituent to divide The analysis method eigenmatrix to obtaining carries out dimension-reduction treatment, as the eigenmatrix of final Emotion identification, uses support vector machine to build Vertical Emotion identification model, weakens temporal characteristics by increasing the natural law of sample in mood model training set, improves mood model Time robustness, thus accurately, objectively carry out Emotion identification.

The method overcoming above-mentioned two problem, both will not reduce the categories of emotions of identification, the data of test set are the most not The foundation of Emotion identification model can be participated in, meet the requirement in reality application.

Embodiment 1

Embodiments provide a kind of emotion electroencephalogramrecognition recognition method improving Emotion identification model time robustness, ginseng Seeing Fig. 1, this emotion electroencephalogramrecognition recognition method comprises the following steps:

101: data acquisition phase collection be different time sections subjects under different emotional states (actively, neutral, disappear Pole) 64 EEG signals led；

102: EEG signals is led in 64 collected and carries out the pretreatment of four steps.Including: it is average that change is referenced to ears； It is downsampled to 500Hz；1-100Hz bandpass filtering；And independent component analysis (independent component Analysis, ICA) go an electrical interference；

103: pretreated EEG signals is used and the algorithm of frequency-division section adaptive tracing can find each user Good separability frequency range, calculate respectively each lead most preferably can the power spectral density of frequency-division section, constitute emotional characteristics matrix；

104: utilize PCA that the emotional characteristics matrix obtained is carried out dimension-reduction treatment, as final feature square Battle array；

105: use support vector machine classifier that the feature in final eigenmatrix is identified, by increasing emotion Model training concentrates the natural law of sample to weaken temporal characteristics, improves the time robustness of mood model, by difference emotional state district Separately, Emotion identification model is set up.

In order to improve the time robustness of Emotion identification model, the embodiment of the present invention is by increasing the sky of sample in training set Number weakens the feature of temporal.Before feature identification, the most respectively the data of every day are normalized.

In sum, the embodiment of the present invention is by finding most preferably can dividing of each user by frequency-division section adaptive tracking algorithm Property frequency range, calculate respectively each lead most preferably can the power spectral density of frequency-division section, constitute emotional characteristics matrix, utilize main constituent The analytic process eigenmatrix to obtaining carries out dimension-reduction treatment, as the eigenmatrix of final Emotion identification, uses support vector machine Set up Emotion identification model, weaken temporal characteristics by increasing the natural law of sample in mood model training set, improve mood model Time robustness, thus accurately, objectively carry out Emotion identification.This invention can be effectively improved Emotion identification model time Between robustness and accuracy.

Embodiment 2

Below in conjunction with Fig. 2, Fig. 3, Fig. 4, Fig. 5, the scheme in embodiment 1 is described in detail, described below:

201: data acquisition phase；

Wherein, brain wave acquisition device is that the 64 of Neuroscan company leads amplifier and Scan 4.5 acquisition system, electrode The standard 10-20 system specified in accordance with international Nao electricity association is placed, and removes leading of 60 conductive electrode outside eye electricity and reference electrode Distribution is as shown in Figure 2.Using right mastoid process as reference electrode during collection, brain forehead top side centre ground connection, all electrodes Impedance is held in below 5k Europe, and sample frequency is 1000Hz.

Every subjects need to carry out 5 secondary data collections in one month, and the time interval between gathering each time is respectively One day, three days, one week and two weeks.Fig. 3 is subjects's experimental period calendar.Every subjects came in the same time of this day Laboratory carries out data acquisition, with video induction three kinds of emotional states tested actively, neutral, passive.

202: data prediction；

EEG signals is led in 64 collected and carries out the pretreatment of four steps.Including: it is average that change is referenced to ears；Fall is adopted Sample goes an electrical interference to 500Hz:1-100Hz bandpass filtering and ICA.

Wherein, reference potential during collection is at auris dextra mastoid process, and this signal amplitude causing You Naoqu to lead is on the low side.Therefore First carry out reference potential conversion, reference potential becomes being positioned at M1, the M2 at mastoid process position, both sides and leads place, it is simple to follow-up data Process.The sample frequency of system is 1000Hz, primarily to meet EEG signals to change rapid requirement.But 1000Hz adopts Sample frequency is far longer than the theoretical sample frequency of Nyquist's theorem, and sample frequency is crossed after conference causes data volume to cross ambassador Continuous treatment effeciency reduces.Therefore, carry out down-sampled to the data collected, the sample frequency of EEG signals is reduced to by 1000Hz 500Hz。

The embodiment of the present invention carries out the bandpass filtering of 1Hz～100Hz to remove DC influence and high-frequency signal.Owing to adopting Collect to EEG signals inevitably contain eye electricity (include eyeball upper and lower, move left and right, blink) and electromyographic signal band The impact come.Wherein eye electricity, the electro-ocular signal especially blinked is strong especially, and being rung maximum by eye film is leading of forehead region Connection.The impact produced for the electro-ocular signal adulterated in EEG signals and electromyographic signal, the embodiment of the present invention passes through isolated component The method that analysis (Independent Component Analysis, ICA) filters is filtered.

Wherein, the method for concrete filtering is known to those skilled in the art, and the embodiment of the present invention is without limitation.

203: can the adaptive tracing method of frequency-division section；

Because different user has different optimal separability frequency ranges, therefore the embodiment of the present invention utilize can frequency-division section adaptive The method that should follow the tracks of (adaptive tracking of discriminative frequency components, ATDFCs) Find the optimum frequency band that different type of emotion can be distinguished, this to extracting feature accurately, to improve classification accuracy rate be the heaviest Want.The calculating process of DFC adaptive tracing method is as shown in Figure 4.

1) Short Time Fourier Transform is used to calculate each time-frequency matrix led；

Therefore, each leads a discrete time-frequency matrix I_n(f,t)。

2) calculate fisher ratio, weigh in class with it and (in same pattern, refer to same feelings in embodiments of the present invention Thread type) and class between the capacity volume variance of (between different mode, referring in embodiments of the present invention between different type of emotion).Its Shown in computational methods such as formula (1), formula (2), formula (3).

$S_w(f,t)=\underset{k=1}{\overset{c}{\Σ}}\underset{n=1}{\overset{n_k}{\Σ}}\left(I_n\right(f,t)-m_k(f,t\left)\right){\left(I_n\right(f,t)-m_k(f,t\left)\right)}^T---\left(1\right)$

$S_B(f,t)=\underset{k=1}{\overset{C}{\Σ}}{n}_k\left(m\right(f,t)-m_k(f,t\left)\right){\left(m\right(f,t)-m_k(f,t\left)\right)}^T---\left(2\right)$

$F_R(f,t)=\frac{S_B(f,t)}{S_w(f,t)}---\left(3\right)$

Wherein, S_w(f,t)、S_B(f,t)、m_k(f, t), m (f, t) and F_R(f t) is two-dimensional matrix.S_w(f,t)、S_B(f, T) represent in class respectively and class inherited, F_R(f t) is fisher ratio, m_k(f t) is the frequent degree of mean time of kth class, m (f, t) is the frequent degree of mean time of all classes, and C represents classification number, C=3, n in the embodiment of the present invention_kIt it is the sample of kth apoplexy due to endogenous wind Number.

3) can fraction weight DW (f) can be tried to achieve by Fisher ratio, its computational methods such as formula (4):

$DW\left(f\right)=\underset{t=1}{\overset{\mathrm{\τ}}{\Σ}}{F}_R(f,t)---\left(4\right)$

Wherein, τ represents time period when STFT calculates.

4) after obtaining DW (f), calculating DFC by wave band iteration selection method, iterations is equal to the band number needing acquisition.

Five steps of Step1 to Step5 cited below can be used to calculate the frequency range of most separability, such as Fig. 5 institute Show.Then weight DW (f) under most separability frequency range is set to zero, then calculates separability and be positioned at deputy frequency range.

Such as, the most frequency range of separability is 9～14Hz, just the DW (f) that 9,10,11,12 is corresponding with 13Hz is set to zero, Calculating second again can frequency-division section；Constantly repeat this process until obtaining the band number needed.

Step1, first determine needs selected frequency range be 1～100Hz, the frequency window of slip between 3～7Hz with step Long 1Hz changes (as shown in Figure 5).Therefore, it can obtain 5 different bandwidth parameters and be designated as BW_j(j=1,2,3,4,5).

Step2, when frequency window moves along the frequency axis of DW (f), calculate Energy distribution α according to formula (5).

$\mathrm{\α}(F_i,{\mathrm{BW}}_j)=\underset{f=F_i-\frac{{\mathrm{BW}}_j}{2}}{\overset{f=F_i+\frac{{\mathrm{BW}}_j}{2}}{\Σ}}DW\left(f\right)---\left(5\right)$

Wherein, F_iRepresent frequency window along frequency axis move time i-th frequency range mid frequency.Such as, frequency window is worked as During a width of 3Hz, then can obtain 97 frequency ranges, such as: 1～4Hz, 2～5Hz, 3～6Hz, 4～7Hz ..., 97Hz～100Hz.

Step3, according to ceiling capacity be distributed α, at all of F_iMiddle selection is optimalSuch as formula (6).

$F_j^{opt}=\arg \underset{F_i}{max}\mathrm{\α}(F_i,{\mathrm{BW}}_j)---\left(6\right)$

To each BW_jOne will be obtainedTherefore, each j, an all corresponding optimal mid frequencyAnd Optimum capacity is distributed

Step4, for relatively each BW_jResolution capability, calculateRelative change, j=(2,3,4,5), utilize public affairs Formula (7) calculates δ_j。

${\mathrm{\δ}}_j=\frac{{\mathrm{\α}}_j^{opt}-{\mathrm{\α}}_{j-1}^{opt}}{{\mathrm{\α}}_j^{opt}}*100\%---\left(7\right)$

Step5, calculate δ_jAfterwards, if a threshold value δ_min。

It is demonstrated experimentally that for different threshold values, such as: 10%, 20%, 30%, 40% ..., threshold value is the least, should Algorithm more levels off to and selects frequency window is the frequency band of 3Hz.

Relatively δ₂And δ_minIf, δ₂More than δ_min, the most then compare δ₃With δ_minSize, until find δ_jLess than δ_min.That The position of j-1 is exactly the frequency range of most separability.

The embodiment of the present invention select each lead first can frequency-division section, second can frequency-division section and the 3rd can the merit of frequency-division section Rate spectrum sets up the eigenmatrix P of every day_Ni*180.Ni is the sample size of i-th day.60 lead * 3 frequency range=180 dimensional features.

P_Ni*180=(P₁,P₂,…,P₁₈₀) (8)

204: PCA dimensionality reduction；

During actual application, there is certain plyability and dependency, if directly between the information contained by parameters They are used for pattern recognition, the overfitting of model parameter can be caused to reduce accuracy and the reliability of classification, and can be because of The speed of classification is reduced for data volume is excessive.Therefore before pattern classification, the embodiment of the present invention first with PCA to each It characteristic vector obtained carries out dimension-reduction treatment.

PCA, according to maximum variance principle, characterizes original data with one group of linear independence and mutually orthogonal new vector The row (or row) of matrix, reaches to compress variable number, rejects redundancy, maximize the purpose preserving effective information.Original to Amount group is (P₁,P₂,…,P₁₈₀), principal component vector group is designated as (F₁,F₂,…,F_m), usual m is less than 180.Then main constituent is with original The relation of Vector Groups is:

$\left\{\begin{array}{c}{F}_1=a_{1,1}{P}_1+a_{1,2}{P}_2+...+a_{1,180}{P}_{180}\\ F_2=a_{2,1}{P}_1+a_{2,2}{P}_2+...+a_{2,180}{P}_{180}\\ ...\\ F_k=a_{k,1}{P}_1+a_{k,2}{P}_2+...+a_{k,180}{P}_{180}\\ ...\\ Fm=a_{m,1}{P}_1+a_{m,2}{P}_2+...+a_{m,180}{P}_{180}\end{array}\right.---\left(9\right)$

Wherein, it is most that F1 contains quantity of information, has maximum variance, referred to as first principal component, F2 ..., Fm successively decreases successively, claims For Second principal component, ..., m main constituent.Therefore the process of principal component analysis is considered as determining that weight coefficient α_k,h(k= 1,…,m；H=1 ... 180) process.

In embodiments of the present invention, Ni the sample (i=1,2,3,4,5) obtained for i-th day, available following matrix represents

$P_{Ni*180}=\left[\begin{array}{cccccc}{P}_{1,1}& P_{1,2}& \mathrm{...}& P_{1,h}& \mathrm{...}& P_{1,180}\\ P_{2,1}& P_{2,2}& \mathrm{...}& P_{2,h}& \mathrm{...}& P_{2,180}\\ .& .& & .& & .\\ .& .& \mathrm{...}& .& \mathrm{...}& .\\ .& .& & .& & .\\ P_{b,1}& P_{b,2}& \mathrm{...}& P_{b,h}& \mathrm{...}& P_{b,180}\\ .& .& .& .& .& .\\ .& .& .& .& .& .\\ .& .& .& .& .& .\\ P_{Ni,1}& P_{Ni,2}& \mathrm{...}& P_{Ni,h}& \mathrm{...}& P_{Ni,180}\end{array}\right]---\left(10\right)$

Wherein, P_b,hIt is the h feature of b sample.

The solution procedure carrying out Feature Dimension Reduction with PCA is as follows:

1) to initial data P_Ni*180Being standardized processing, the element in matrix deducts the average of column, then divided by The standard deviation of column so that the average of each variable is 0, and variance is 1, obtains matrix P_Ni*180 ^*。

P_Ni*180 ^*=[y_b,h]_Ni*180, b=1,2 ..., Ni；H=1,2 ..., 180 (11)

$y_{b,h}=\[P_{b,h}-\stackrel{\‾}{P_h}\]/S_h---\left(12\right)$

Wherein,

2) its covariance matrix C is then sought_180*180,P_Ni*180 ^*In wantonly two row between can calculate the covariance between two variablees, Then covariance matrix is obtained:

3) to covariance matrix C_180*180Carry out characteristic root decomposition, obtain characteristic root matrix Λ_180*180And characteristic vector U_180*180。

C_180*180=U_180*180Λ_180*180U_180*180(14) wherein, characteristic vector U_180*180As the coordinate axes of main constituent, Constitute new vector space,Wherein, characteristic root λ_r(r=1,2 ..., 180) size represent The quantity of information that the r main constituent is contained.U_180*180' it is U_180*180Transposed matrix.

4) initial data P is sought_Ni*180Projection in new vector space, i.e. principal component vector group F_Ni*180:

F_Ni*180=P_Ni*180U_180*180 (15)

5) accumulation contribution rate is sought.The characteristic root size of each main constituent represents its number containing quantity of information.Ask front k master The accumulation contribution rate of composition (k=1 ..., 180).

${\mathrm{pre}}_k=\frac{\underset{i=1}{\overset{k}{\Σ}}{\mathrm{\λ}}_i}{\underset{i=1}{\overset{180}{\Σ}}{\mathrm{\λ}}_i}---\left(16\right)$

Wherein, λ i is the ith feature root obtained.

6) the selected accumulation contribution rate preset, makes front d main constituent F_Ni*dAs new data carry out pattern recognition (d < 180)。

Such as: draw altogether 7 main constituents.The contribution rate of first main constituent F1 is 48%, and the contribution rate of F2 is 32%, The contribution rate of F3 is 15%, and F4, F5, F6, F7 contribution rate altogether is 5% (contribution rate of 7 main constituents comes to 100%). So the accumulation contribution rate of first three main constituent (F1, F2, F3) is to 95%, say, that first three main constituent has contained 7 The information of individual main constituent 95%, then, select these three main constituent to carry out pattern recognition as new data, in guarantee information amount While reduce the dimension of eigenmatrix.

205: the foundation of Emotion identification model；

After Feature Dimension Reduction, use support vector machine (Support Vector Machine, SVM)^[3]Set up Emotion identification mould Type, identifies the emotional state that user is current.In the pattern recognition stage, a part of sample is used for setting up grader, referred to as training set, Remaining sample is used for testing classification device, referred to as test set.

The embodiment of the present invention first passes through increases in training set the natural law of sample to improve the time Shandong of Emotion identification model Rod.So in the pattern recognition stage, the data of four days being used for training pattern, the sample remaining a day is attributed to test set.This Sample is conducive to training set to extract the feature relevant to emotion, and weakens the feature of temporal.

Before setting up grader, it is necessary first to the data of every day are carried out row normalization respectively, normalizes to [-1.1], Obtain eigenmatrix PP_Ni*d, (d is the dimension of eigenmatrix after dimensionality reduction)

PP_i,j=(y_max-y_min)*(F_i,j-F^j _min)/(F^j _max-F^j _min)+y_min,

Wherein, y_max=1, y_min=-1；For the eigenmatrix F after dimensionality reduction_Ni*dJth row minima, in like manner,It is characterized matrix F_Ni*dJth row maximum.After normalization, SVM classifier is utilized to set up Emotion identification model.

In sum, the embodiment of the present invention is by frequency-division section adaptive tracing method searching out most preferably can dividing of each user Frequency range；By PCA, eigenmatrix is carried out Feature Dimension Reduction；Weaken by improving the natural law of sample in training set Temporal feature, accurate stable carries out Emotion identification in real time then.This invention can be effectively improved Emotion identification mould The time robustness of type and accuracy.

Embodiment 3

Below in conjunction with Fig. 6, the scheme of embodiment 1 and 2 is carried out feasibility checking, described below:

Fig. 6 is the recognition correct rate under 9 tested different training sample natural law.Transverse axis is the natural law N of sample in training set (N=1,2,3,4), the sample training of i.e. N days, the sample of remaining 5-N days is tested.Under the conditions of the longitudinal axis is N days obtained Mean accurate rate of recognition.By in Fig. 6 it can be seen that along with the increase of natural law in training set, accuracy improves, accuracy and instruction The natural law practicing sample is proportionate；The sample of 4 days is for training the grader sample training compared to 1 day, and its accuracy carries High by about 10%, and there is significant difference (p < 0.01).This method also demonstrating present invention proposition is effective.

The embodiment of the present invention by can frequency-division section adaptive tracing method find each user most preferably can frequency-division section, main constituent Analytic process carries out Feature Dimension Reduction, and increases in Emotion identification model the natural law of sample in training set and enhance the spy that emotion is relevant Levy, weaken the feature of temporal, thus improve the time robustness of Emotion identification model.This Fig. 6 illustrates to increase In training set, the natural law of sample can significantly improve the time robustness of grader.The present invention can be effectively improved Emotion identification mould The time robustness of type and preparatory, moves towards application for Emotion identification from laboratory and provides technical support.

List of references

[1]PICARD R W,VYZAS E,HEALEY J.Toward machine emotional intelligence: Analysis of affective physiological state[J].Pattern Analysis and Machine Intelligence,IEEE Transactions on,2001,23(10):1175-91.

[2]CHUEH T-H,CHEN T-B,LU H H-S,et al.Statistical Prediction of Emotional States by Physiological Signals with Manova and Machine Learning [J].International Journal of Pattern Recognition and Artificial Intelligence, 2012,26(04):

[3]HIDALGO-MU OZ AR,L PEZ M M,SANTOS I M,et al.Application of SVM-RFE on EEG signals for detecting the most relevant scalp regions linked to affective valence processing[J].Expert Systems with Applications,2013,40(6): 2102–8.

It will be appreciated by those skilled in the art that accompanying drawing is the schematic diagram of a preferred embodiment, the invention described above embodiment Sequence number, just to describing, does not represent the quality of embodiment.

The foregoing is only presently preferred embodiments of the present invention, not in order to limit the present invention, all spirit in the present invention and Within principle, any modification, equivalent substitution and improvement etc. made, should be included within the scope of the present invention.

Claims (6)

1. the emotion electroencephalogramrecognition recognition method improving Emotion identification model time robustness, it is characterised in that described emotion brain Electricity recognition methods comprises the following steps:

EEG signals is led in 64 collected and carries out pretreatment, including: it is average that change is referenced to ears；It is downsampled to 500Hz；1- 100Hz bandpass filtering；And utilize the algorithm of independent component analysis to remove eye electrical interference；

Pretreated EEG signals is used and the algorithm of frequency-division section adaptive tracing can find the optimal separability of each user Frequency range, calculate respectively each lead most preferably can the power spectral density of frequency-division section, constitute emotional characteristics matrix；

PCA is utilized the emotional characteristics matrix obtained to be carried out dimension-reduction treatment, as final eigenmatrix；

Use support vector machine classifier that the feature in final eigenmatrix is identified, train by increasing mood model The natural law concentrating sample weakens temporal characteristics, improves the time robustness of mood model, is distinguished by difference emotional state, sets up Emotion identification model.

A kind of emotion electroencephalogramrecognition recognition method improving Emotion identification model time robustness the most according to claim 1, its Being characterised by, described method also includes:

Gather different time sections subjects 64 EEG signals led under different emotional states.

A kind of emotion electroencephalogramrecognition recognition method improving Emotion identification model time robustness the most according to claim 1, its Being characterised by, described pretreated EEG signals employing the algorithm of frequency-division section adaptive tracing can find each user The step of good separability frequency range particularly as follows:

1) Short Time Fourier Transform is used to calculate each time-frequency matrix led；

2) fisher ratio, the capacity volume variance in weighing class with it and between class are calculated；

3) can fraction weight DW (f) can be tried to achieve by Fisher ratio；Calculating DFC by wave band iteration selection method, iterations is equal to Need the band number obtained；Optimal separability frequency range is obtained from the band number obtained.

A kind of emotion electroencephalogramrecognition recognition method improving Emotion identification model time robustness the most according to claim 1, its Be characterised by, described from the band number obtained, obtain optimal separability frequency range particularly as follows:

When frequency window moves along the frequency axis of DW (f), calculate Energy distribution；It is distributed α, all of according to ceiling capacity Band number select optimalCalculateRelatively change δ_j；

If a threshold value δ_min, compare δ₂And δ_minIf, δ₂More than δ_min, the most then compare δ₃With δ_minSize, until find δ_jLess than δ_min, then the position of j-1 is exactly the frequency range of optimal separability.

A kind of emotion electroencephalogramrecognition recognition method improving Emotion identification model time robustness the most according to claim 1, its It is characterised by, described utilizes PCA the emotional characteristics matrix obtained to be carried out dimension-reduction treatment, as final feature The step of matrix particularly as follows:

1) it is standardized initial data processing, obtains original matrix；Then its covariance matrix is sought；Covariance matrix is entered Row characteristic root decomposes, and obtains characteristic root matrix and characteristic vector；

2) original matrix projection in new vector space, i.e. principal component vector group are asked；

3) the characteristic root size of each main constituent represents its number containing quantity of information, seeks the accumulation contribution rate of front k main constituent；

4) the selected accumulation contribution rate preset, makes front d main constituent F_Ni*dPattern recognition is carried out as new data.

A kind of emotion electroencephalogramrecognition recognition method improving Emotion identification model time robustness the most according to claim 1, its Being characterised by, the feature in final eigenmatrix is identified by described use support vector machine classifier, by increasing feelings Thread model training concentrates the natural law of sample to weaken temporal characteristics, improves the time robustness of mood model, by difference emotional state Distinguish, set up the step of Emotion identification model particularly as follows:

The data of every day are carried out row normalization respectively, normalizes to [-1.1], obtain eigenmatrix；

SVM classifier is utilized to set up Emotion identification model；In modeling process, the data of many days are put into training set, divide to improve The time robustness of class device.