CN102622527A - Taboo searching method for selection of galvanic skin response signal features - Google Patents

Abstract

The invention discloses a taboo searching method for selection of galvanic skin response signal features, which includes the following steps: sequence backward algorithm is adopted to form N-1 rows and a two-dimensional table L of N lines, wherein the N represents total number of dimensions of selected features, each line represents one feature, each row is called one space, an nth space selects n features, and n is larger than or equal to 1 and less than or equal to N-1; a value of each element in the table is expressed by 0 or 1, 0 represents that the element is not selected when feature selection is performed, and 1 represents that the element is selected; the selected features in each space are solved by adopting taboo searching algorithm to obtain a table S formed by solution of each space; and a feature with the largest fitness function in each space serves as a final feature selection result. The taboo searching method for the selection of galvanic skin response signal features not only obtains effective features of skin electric signals with large contribution to emotion recognition, but also overcomes the shortcoming that a basic taboo searching algorithm is easy to get into local optimum.

Description

A kind of taboo search method that is used for the selection of electrodermal response signal characteristic

Technical field

The present invention relates to a kind of physiological signal emotion identification method, relate in particular to the feature selecting of electrodermal response signal.

Background technology

Emotion recognition is an important component part in the emotion computing field, if let computing machine have emotion, at first will let it can discern emotion, and it is particularly important that emotion recognition seems.Emotion recognition (Emotion Recognition) is that the prerequisite environment that expression, behavior and the emotion through the observer produces is inferred affective state.Have only and regard emotion recognition as a kind of pattern recognition problem, emotional expression is seen the operation mode composition problem, and computing machine carries out emotion communication and just has feasibility.The object of emotion recognition research mainly contains five kinds of human face expression, speech intonation, human posture, text (on the psychology commonly used questionnaire method) and physiological signals.Preceding four kinds of research objects are more directly perceived, all are that mode with health and behavior shows voluntarily or the complex patterns of signal voluntarily not, can't observe potential affective state.And physiological change is not controlled by people's subjectivity; It shows by health is objective; Thereby adopt the data of physiological signal gained more can react the true emotion of being tried objectively, it more has robustness and objectivity, but the research that is based on the physiological signal emotion recognition is difficulty.

Before the present invention; " Affective Pattern Classification.Perceptual Computing Section Technical Report (Elias Vyzas and Rosalind W.Picard; 1998; 473) " is middle to prove that it is effective adopting the method that is characterized as the emotion recognition carrier of physiological signal for the first time to the breadboard Picard professor of U.S. MIT for its laboratory gives a technical report in 1998; And discern 8 kinds of different emotions states of performer through extracting beat 40 signal characteristics of (BVP) four kinds of physiological signals of electrodermal response signal (GSR), breath signal (RSP), electromyographic signal (EMG) and blood volume; Adopt SFFS (sequence forward direction selection algorithm) and these two kinds of feature selection approachs of Fisher, and to have selected multiple sorter, the discrimination that obtains be 81.25%; " Emotion Recognition Using Physiological and Speech Signal in Short-Term Observation (Jonghwa Kim and Elisabeth Andr.2006; LNAI 4021; pp.53-64) " in; It is 55% that Jonghwa takes the discrimination of 4 kinds of extreme moods of method identification that physiological signal combines with voice signal, and in this article, the author has extracted physiological signal and voice signal totally 138 characteristics; And carrying out feature selecting with SBS (after the sequence to selection algorithm), used sorter is LDA (a linear discrimination classification device); In " Emotion Recognition Using Physiological Signals (Lan Li and Ji-hua Chen; ICAT 2006.LNCS 4282; 2006:437～446) "; When 22 characteristics that people such as Lan Li have extracted GSR, SKT, RSP and four kinds of physiological signals of ECG were discerned these three kinds of states of happiness, fear and calmness (not having any mood), discrimination was up to 86.7%.In feature selecting, the researcher adopts traditional feature selection approach, have in addition do not have feature selecting, directly the characteristic of extracting is used for the emotion classification; The selection of sorter is also different, all is rule of thumb to select to think the sorter of good classification effect.Because the physiological signal that each researcher adopts and the difference of affective style cause recognition result different.In the physiological signal that the researcher adopts, the GSR signal is all arranged, the GSR characteristic that they extract, normally several statistical natures such as the first order difference on the time domain, average do not have very big contribution but sum up concrete which characteristic at last to the identification emotion.

Summary of the invention

This object of the invention provides the high taboo search method that is used for the selection of electrodermal response signal characteristic of a kind of effective recognition rate.

To achieve these goals, adopt following technical scheme: a kind of taboo search method that is used for the selection of electrodermal response signal characteristic, it is characterized in that: said method comprises the steps:

To algorithm, it is capable to form a N-1 after the employing sequence, the bivariate table L of N row, and wherein N is the total dimension of characteristic of selection, and each row is represented a characteristic, and each is gone and is called a space, and wherein n characteristic, 1≤n≤N-1 are selected to have in n space; The value of each element is with " 0 " or " 1 " expression in the table, and " 0 " representative this element when carrying out feature selecting does not have selected, and " 1 " represents this element selected;

Adopt tabu search algorithm to find the solution to the characteristic of choosing in each space, obtain the table S that separates composition in each space;

Select the maximum final feature selecting result of conduct of fitness function in each space.

Wherein the step of tabu search algorithm is in each space:

S1: initially establish taboo table T=Φ, taboo length is set, the greatest iteration step number is set; With in this space through the value that obtains to algorithm after the sequence as initial solution; And it as the tentative globally optimal solution Bestsofar in this space and the starting point of iterative search; Be current locally optimal solution cand, calculate fitness function

value of initial solution; Wherein ratA and ratB represent the correct recognition rata of target emotion and the correct recognition rata of non-target emotion respectively.

S2: judge whether to satisfy the stopping criterion of greatest iteration step number,, will put into table S to Bestsofar, finish this SPATIAL CALCULATION as satisfying the calculating that then stops this space;

As do not satisfy, then be the iteration starting point of next time with current locally optimal solution cand;

S3: generation N is individual to have candidate's disaggregation that same characteristic features is selected number with this space;

S4: optimizing: calculate the value of the fitness function f of each candidate solution, concentrate from candidate solution and select maximum the separating of fitness function value,

With this separate with the taboo table in separate comparison; If this is separated not in the taboo table, with the value of wherein bigger value as current locally optimal solution and Bestsofar; Then change S5;

If this is separated in the taboo table, and do not satisfy the special pardon criterion, with second largest the separating of fitness function value as locally optimal solution;

If this is separated in the taboo table, and satisfy the special pardon criterion, in the taboo table this separated in advance discharged, and as locally optimal solution, and with the fitness function value comparison of this fitness function value of separating and Bestsofar, as greater than, then separate as Bestsofar with this; Wherein specially pardon criterion and refer to, certain value in the taboo table stipulated number occurs as locally optimal solution in iterative process;

S5: upgrade the taboo table: current locally optimal solution is write the taboo table, change S2.

The method that generates candidate's disaggregation in the said S3 step is: i position cand (i) value with cand becomes 1-cand (i) respectively, searches since the next bit of i position, runs into the p position that value equals 1-cand (i), its value is become 1-cand (p) back finish; If P=N, then P continues to look into since 1, runs into the p position that value equals 1-cand (i), its value is become 1-cand (p) back finish;

1≤i≤N wherein, 1≤P≤N, N are the total dimension of characteristic.

The present invention adopts improved tabu search algorithm combination Fisher sorter completion selection course in the intelligent optimization algorithm; Not only drawn validity feature, and overcome the shortcoming that basic tabu search algorithm is absorbed in local optimum easily the bigger skin electric signal of emotion recognition contribution.Experiment simulation is the result show, the emotion recognition field that improved intelligent optimization algorithm is used for physiological signal is fully feasible.

Figure of description

Fig. 1 is an intermediate data distribution plan in the Fisher sorter;

Fig. 2 is a final test data profile in the Fisher sorter.

Embodiment

Below in conjunction with accompanying drawing and embodiment the present invention is done further elaboration.

A kind of emotion recognition feature selection approach based on the electrodermal response signal comprises the following steps:

1. emotion excites the foundation of material database

Effectively excite the emotion of being tried, obtaining effective emotion physiological signal is the precondition of doing follow-up study, so the emotion of choosing excites material very important.

(1) neutral material

The mode that adopts light music and scenery picture to combine is made seven neutral materials; Long 2 minutes of each neutral material; Purpose is to let the mood of being tried before exciting emotion, be in tranquility, and the physiological signal that the quilt that writes down this moment tries is tried the benchmark of other emotion excitation signals as this.

(2) emotion material

Hobby and interest to the people of China youth; Select meticulously can bring out respectively the happiness that tried, surprised, detest, the Chinese movie or television of the high television ratings of sad, indignation and frightened six kinds of emotions is acute; Montage goes out more than 100 fragment and excites video alternative as emotion; Then through many people, repeatedly repeated validation its arouse the reliability of emotion; Finally therefrom determine corresponding six emotions and excite material, and excite material to begin to add before the background note of suitable literal as material in each emotion.Each emotion material was about 5 minutes; Wherein, glad emotion excites material to come from the fragment montage of " martial arts circles's unofficial history ", comes from " magic " in surprise; Detest and come from " kermes snow "; Sadness comes from " warm spring ", and indignation comes from " Tokyo trial ", and fear comes from " The Ring ".

(3) material is integrated

7 neutral materials are positioned over six emotion materials begin before and end, regulate, become a cover experiment material according to glad, surprised, detest, sadness, indignation, frightened sequence integration as transition.The material structural order is: experimental instruction (2 minutes → neutral material 1 (2 minutes) → material 1 background introduction (30 seconds) → emotion excite material 1 (about 5 minutes) → 1 instruction that fills in questionnaires (30 seconds) → neutral material 2 (2 minutes) → material 2 background introductions (30 seconds) → ... → 6 instructions that fill in questionnaires (30 seconds), about 50 minutes of a whole set of material.

2.GSR the affection data collection of signal

The polygraph MP150 that experiment adopts U.S. Biopac company to produce gathers by the original GSR signal of examination under each emotion (comprising neutrality) excited state.Because the useful signal frequency range of GSR mainly concentrates on below the 0.2Hz, according to nyquist sampling theorem, the SF in the gatherer process is made as 20Hz.Connect instrument and adopt TSD203 finger electrode, in its groove, inject conductive paste, touch on the body that is tried with non-invasive mode.

Also will be in the experiment to provided a mood subjective account by examination; Have in the account seven kinds of moods (tranquil, glad, surprised, detest, sad, indignation, fear) and mood intensity (1-very a little less than, 1-a little less than, 3-is general; 4-is stronger; 5-is very strong), after each emotion excited the material broadcast to finish, material was pointed out automatically to let and is filled in own at this moment real mood and strong and weak degree thereof by examination.According to the data that the emotion questionnaire screening of being filled in by examination is effectively aroused, the GSR signal of 80 seconds length of each emotion intercepting amounts to 342 data samples as the raw data of after-stage.

3. the pre-service of raw data

Possibly there are two problems for the raw data that obtains: the one, possibly be mingled with the noise or the intrinsic noise of MP150 machine itself of other physiological signals; The 2nd, owing to there is individual difference, the different emotion standards of being tried are different, and are inconsistent to the GSR reaction signal of same material.To above two problems, to raw data carry out smoothly, operation such as filtering and noise reduction, removal baseline, the emotion signal of each sample is done standardization with respect to disposition sense signal wherein then.

At first, the electrodermal response signal that collects is faint bioelectrical signals, and the interference in the gatherer process mainly contains baseline wander, other physiological signal interference, electrode contact noise, electromagnetic interference (EMI) and motion artifact etc.Because the useful frequency band of electrodermal response signal is mainly below 0.2Hz; The frequency band of its undesired signal and electrodermal response signal is nonoverlapping; Therefore can adopt Butterworth filter filtering out-of-band noise, remove high frequency interference; For the situation that signal is flooded by noise fully, then intercepting abandons these nugatory data.

Next, because the testee who recruits is different, its GSR signal exists individual difference, this otherness is embodied in different people and is facing identical environment at the same time, and same individual faces under the varying environment at different time.For set up general emotion recognition system based on the GSR signal, must remove this individual difference after, just the model that obtains possibly promoted, so need do standardization to data, rule as follows:

D _o＝D _emotion-D _calm (1)

Wherein, D _EmotionBe the original affection data behind the noise reduction, D _CalmBe the same GSR response data that under neutral montage, is write down by examination, the D that obtains ₀Be through the data after the standardization.

During concrete operations, because former data length is 80s, SF is 20Hz, so be 1600 stored in form data according to number of data points, then only need subtract each other each some correspondence of affection data and these tranquil data of being tried and can obtain normal data.

4.GSR the affective characteristics of signal extracts and handles

Under the different affective states, skin electricity level value is difference to some extent, and amplitude of variation also has difference, and particularly frightened emotion is the most obvious, secondly is glad.From visually being easy to pick out of the variation of electrodermal response signal, therefore need not carry out complicated more analysis with emotion.To GSR signal own characteristic; Directly from time domain and frequency-region signal; And time-domain signal carried out extracting on the signal after first order difference, second order difference are calculated calculate some statistical value and form primitive character set; Comprise poor, the minimum value ratio and the maximal value ratio of average, intermediate value, standard deviation, minimum value, maximal value, maximin, amount to 30 characteristics.

Wherein, first order difference can embody the variation tendency and variation speed of signal, if first order difference has been got absolute value, has just ignored the trend that changes, and has only considered to change speed.First order difference can be used for the local extreme point of detection signal, and second order difference can be used for the local flex point of detection signal.

From 30 characteristics of GSR signal extraction, 24 of temporal signatures, 6 of frequency domain characters.The characteristic of being extracted constitutes by " Name-english abbreviation ", as: sc1Diff-median, sc2Diff-std, sc-mean, sc-minRatio, scfft-mean, sc2gDiff-mean, sc2aDiff-mean etc.The detailed features that extracts is as shown in table 1:

The statistical nature of table 1GSR signal is described

Wherein, X _nBe used for n sample of expression signal;

N representes data length to be analyzed;

Max representes the maximal value of data;

Min representes minimum value.

u _xAnd σ _xBe respectively the average and the standard deviation of sample.

With frightened emotion is example, and 30 concrete numerical value of characteristic that extract some electrodermal response signals that is tried are as shown in table 2.

30 concrete eigenwerts that a certain GSR signal that is tried of table 2 extracts under frightened emotion

Following formula is to the normalization processing of sample and the extraction formula of statistical nature.

For making things convenient for the relatively unified of data, all characteristics are carried out numerical value normalization handle.

${\tilde{X}}_i=\frac{X_{i0}-X_{i\min }}{X_{i\max }-X_{i\min }}---\left(2\right)$

X in the formula _I0Represent the eigenwert of certain sample of the individual characteristic of i (i ∈ (0,30)), X _Imax, X _IminRepresent the maximal value and the minimum value of these dimensional feature data respectively,

Be the result after handling through normalization, 30 all values of dimensional feature data that obtain like this are all between 0-1.According to the result of table 2, the result after normalization is handled is as shown in table 3.

Result after table 3 normalization is handled

sc-mean	sc-median	sc-std	sc-min	sc-max	sc_range
						0.6579	0.6572	0.0078	0.7110	0.5897	0.0065
sc-minRatio	sc-maxRatio	sc1Diff-mean	sc1Diff-median	sc1Diff-std	sc1Diff-min
						0.1944	0.1944	0.2641	0.7834	0.0063	0.9972
sc1Diff-max	sc1Diff?range	sc1Diff-minRatio	sc1Diff-maxRatio	sc1gDiff-mean	sc2Diff-std
						0.0044	0.0042	0.2367	0.2415	0.2565	0.0026
sc2Diff-min	sc2Diff-max	sc2Diff?range	sc2Diff-minRatio	sc2Diff-maxRatio	sc2gDiff-mean
						0.9975	0.0018	0.0021	0.5382	0.5382	0.1013
scfft-mean	scfft-median	scfft-std	scfft-min	scfft-max	scfft_range
						0.0049	0.0031	0.0066	0.0035	0.0051	0.0055

With characteristic " sc-median " is example explanation computation process.The sc-median eigenwert that the GSR signal that this is tried extracts under frightened emotion is 0.0680; And this characteristic is 5.0597 at 342 samples by the maximal value in trying; Minimum value is-9.5025; So according to formula (2) calculating (0.0680-(9.5025))/(5.0597-(9.5025)) are arranged, the value that promptly obtains in the table 3 is 0.6572.

5. adopt improved tabu search algorithm to combine the Fisher sorter to carry out the selection and the classification of affective characteristics subclass

TABU search (Tabu Search is called for short TS) is a kind of simulation to human thinking's process itself, and it reaches through the taboo (also can be described as memory) to some locally optimal solutions admits relatively poor the separating of a part, thereby reaches the purpose of jumping out Local Search.Its most important thought is to have adopted the taboo technology; Simultaneously in order not miss " the moving " that produces optimum solution as far as possible; " despising criterion " strategy is also adopted in TABU search, and then can ignore its taboo attribute and still adopt it is current selection, to avoid losing good state.

This algorithm is compared with traditional optimized Algorithm has the ability of well climbing the mountain and regional centralized search and the better balance of overall scatter searching ability, for overcoming the deficiency of rudimentary algorithm, has proposed the emotion recognition that a kind of improved tabu search algorithm is applied to electromyographic signal.This algorithm is divided into a plurality of subspaces with the search volume, utilizes after the sequence to what algorithm obtained in each subspace and separates the initial solution as TABU search, divides time-like; Select the linear discriminant function Fisher sorter that writes down in " image model identification (Yang Shuying, publishing house of Tsing-Hua University, 2006; 95-101) " book for use, fitness function is defined as the classification correct recognition rata, and regards one type of emotion that needs branch away as the target emotion; Regard all the other five types of emotions as non-target emotion, as classifying with other five types of emotions to glad, then happiness is the target emotion; Other five types of emotions are non-target emotion, begin search in the subspace of correspondence from each initial solution then, obtain the optimal feature subset of each subspace; At last these are separated and estimate, find out the optimal feature subset in whole space.

5.1 fitness function

The foundation of the quality of character subset that evaluation obtains combination is two types of correct recognition ratas selecting for use the linear discriminant function Fisher classifier calculated that writes down in " image model identification (Yang Shuying; publishing house of Tsing-Hua University; 2006,95-101) " book to go out, fitness function is defined as follows:

$f=\sqrt{({\mathrm{ratA}}^2+{\mathrm{ratB}}^2)}$

Wherein establish ratA and ratB and represent the correct recognition rata of target emotion and the correct recognition rata of non-target emotion respectively.

Illustrate ratA and the computation process of ratB value in the Fisher linear classifier below.

The ultimate principle of Fisher classification finds an only axis of projection exactly; Make the class spacing of two types of samples on this big as far as possible, spacing is as far as possible little in the sample class, can reach best classifying quality like this; How to find this axis of projection, the matter of utmost importance that will solve exactly.

Step 1: the sample data collection that comprises target and non-target emotion is set, is used to obtain optimum axis of projection.It is following that data are set:

The target emotion comprises 4 samples and 2 characteristics, and data are following:

$w1=\left[\begin{array}{cc}16.69& 26.66\\ 27.95& 6.07\\ 12.15& 6.99\\ 7.36& 13.94\end{array}\right]$

Non-target emotion comprises 7 samples and 2 characteristics, and data are following:

$w2=\left[\begin{array}{cc}-9.88& -24.14\\ 10.58& -6.94\\ 2.22& -13.91\\ -7.93& 3.29\\ -3.88& 5.98\\ -4.86& 1.47\\ 4.28& -1.01\end{array}\right]$

Step 2: calculate all kinds of sample average vector m _i, m wherein _iBe the average of all kinds of samples, N _iBe ω _iThe number of samples of class:

$m_i={\left(\frac{1}{N_i}\underset{X\∈w_i}{\mathrm{\Σ}}X\right)}^T$

(16.69+27.95+12.15+7.36)/4＝16.0375；

Calculate

$m_1=\left[\begin{array}{c}16.0375\\ 13.4150\end{array}\right]$

$m_2=\left[\begin{array}{c}-1.3529\\ -5.0371\end{array}\right]$

Step 3: calculate sample within class scatter matrix S _iAnd dispersion matrix S between total type _w:

$S_i=\underset{X\∈{\mathrm{\ω}}_i}{\mathrm{\Σ}}(X-m_i){(X-m_i)}^T,i=\mathrm{1,2}$

S _w＝S ₁+S ₂

16.69-16.0375＝0.6525；26.66-13.4150＝13.245

$S_1=\left[\begin{array}{cccc}0.6525& 11.9125& -3.8875& -8.6775\\ 13.245& -7.3450& -6.4250& 0.5250\end{array}\right]\×\left[\begin{array}{cc}0.6525& 13.2450\\ 11.9125& -7.345\\ -3.8875& -6.425\\ -8.6775& 0.525\end{array}\right]=\left[\begin{array}{cc}232.7451& -58.4334\\ -58.4334& 270.9353\end{array}\right]$

$S_2=\left[\begin{array}{cc}321.5449& 25.7370\\ 25.7370& 696.5471\end{array}\right]$

$S_w=\left[\begin{array}{cc}554.2900& -32.6965\\ -32.6965& 967.4824\end{array}\right]$

Step 4: the calculating optimum vector $w=S_w^{-1}(m1-m2);$

$w^{*}={\left[\begin{array}{cc}554.2900& -32.6965\\ -32.6965& 967.4824\end{array}\right]}^{-1}\left(\left[\begin{array}{c}16.0375+1.3529\\ 13.4150+5.0371\end{array}\right]\right)=\left[\begin{array}{c}0.0326\\ 0.0202\end{array}\right]$

Step 5: calculated threshold y ₀:

If 2 dimension samples project in the one dimension Y space, its average is so:

${\tilde{m}}_i=\frac{1}{N_i}\underset{y\∈{\mathrm{\ζ}}_i}{\mathrm{\Σ}}Y=\frac{1}{N_i}\underset{x\∈w_i}{\mathrm{\Σ}}{w}^{T}x=w^T\left(\frac{1}{N_i}\underset{x\∈w_i}{\mathrm{\Σ}}x\right)=w^T{m}_i$

$\mathrm{<figure-callout\; id="0"\; label="y"\; filenames="120413144441.png,120413144444.png"\; state="\{\{state\}\}">y</figure-callout>}0=\frac{{\tilde{m}}_1}{2}+\frac{{\tilde{m}}_2}{2}$

${\tilde{m}}_1=\left[\begin{array}{cc}0.0326& 0.0202\end{array}\right]\&times;\left[\begin{array}{c}16.0375\\ 13.4150\end{array}\right]=0.7929$

${\tilde{m}}_2=-0.1457$

$\mathrm{<figure-callout\; id="0"\; label="y"\; filenames="120413144441.png,120413144444.png"\; state="\{\{state\}\}">y</figure-callout>}0=\frac{0.7929}{2}+\frac{-0.1457}{2}=0.3236$

Step 6: the vector of getting maximum value of drawing is separated as shown in Figure 1:

X=[40:0.1:40] is the X axle

Y=x*0.0202/0.0326 is the Y axle

The axis of projection of straight line for obtaining, the data of " " expression target emotion, the data of the non-target emotion of " * " expression.

Step 7: provide test data, test the target emotion of this model and the correct recognition rata ratA and the ratB value of non-target emotion:

The target emotion comprises 4 samples and 2 characteristics, and data are following:

$p1=\left[\begin{array}{cc}-8.5714& -11.6228\\ 14.1014& 2.5585\\ 7.4654& 8.6988\\ 14.4700& 13.9620\end{array}\right]$

Non-target emotion comprises 6 samples and 2 characteristics, and data are following:

$p2=\left[\begin{array}{cc}21.4747& 6.5058\\ 3.9631& 4.7515\\ -2.6728& -9.4298\\ -10.4147& -0.3655\\ -14.2857& -3.8743\\ -14.8387& -6.7982\end{array}\right]$

The test data distribution plan is as shown in Figure 2:

" " represents threshold value, cuts apart two types of emotions, and projection is the target emotion greater than threshold value; Projection is non-target emotion less than threshold value, and " zero " represents the target emotion, and " ☆ " represents non-target emotion; From figure, can find out in 4 target emotion samples have 1 to be identified as non-target emotion, so

$\mathrm{ratA}=\frac{3}{4}\&times;100\%=75\%$

There are 2 samples to be identified as the target emotion in 6 non-target emotion samples, so

$\mathrm{ratB}=\frac{4}{6}\&times;100\%=66.67\%$

5.2 it is following that improved tabu search algorithm carries out the step of feature selecting:

Step 1: will move " the feature selecting algorithm Review Study (and hair is brave; Zhou Xiaobo. the feature selecting algorithm Review Study. pattern-recognition and artificial intelligence [J]; 2007,4:V01.20 No.2) " in the literary composition after the recorded sequences to algorithm, combine with the Fisher sorter to algorithm after the employing sequence; After calculating; Each element among the result who obtains with " 0 " or " 1 " expression (" 0 " representative this element when carrying out feature selecting does not have selected, and " 1 " represents this element selected), and is placed on this result among the table L; J representes the line number of L, establishes j=1, is first row of table L;

Step 2: initialization: taboo table T=Φ, taboo length is set, and the greatest iteration step number is set; Initial solution x (j)=L (j; :) (the initial solution in capable j the space as tabu search algorithm of j of table L); And, calculate the fitness function of initial solution its starting point (current locally optimal solution (cand)) as the tentative globally optimal solution Bestsofar and the iterative search in this space $f=\sqrt{({\mathrm{RatA}}^2+{\mathrm{RatB}}^2)}$ Value.

Step 3: the end condition of single search volume: the stopping criterion that judges whether to satisfy the greatest iteration step number; As satisfying the calculating that then stops this space; And it is capable to be placed on the j that shows S to Bestsofar; Changeing Step 7 then, as not satisfying, then is the iteration starting point of next time with current locally optimal solution (cand);

Step 4: generate N candidate's disaggregation: respectively with i (1≤i≤N of cand; N is the total dimension of characteristic) position cand (i) value becomes 1-cand (i); Search since the next bit of i position, run into the p position that value equals 1-cand (i), its value is become 1-cand (p) back finish; If P=N, then P continues to search since 1, runs into the p position that value equals 1-cand (i), its value is become 1-cand (p) back finish;

Step 5: optimizing: calculate the value of the fitness function f of each candidate solution, concentrate from candidate solution and select maximum the separating of fitness function value, with this separate with the taboo table in separate comparison; If this is separated in the taboo table,, separate as locally optimal solution relatively poor if do not satisfy the special pardon criterion; If satisfy the special pardon criterion; In the taboo table this separated in advance discharge, and as locally optimal solution, and with the fitness function value comparison of this fitness function value of separating and Bestsofar; As greater than, then separate as Bestsofar with this; Wherein specially pardon criterion and refer to, certain value in the taboo table stipulated number occurs as locally optimal solution in iterative process;

If this is separated not in the taboo table, with the value of wherein bigger value as current locally optimal solution and Bestsofar;

Step 6: upgrade the taboo table: current locally optimal solution is write the taboo table, change Step 3;

Step 7: forward next space search: j=j+1 to; Total dimension changes Step 2 if j is not equal to characteristic;

Step 8: compare separating of each space: select corresponding the separating as optimum solution (Y) of Max (f) in the S table, the output optimum solution.

Only comprising 4 feature selection processes with primitive character below is that example is specifically introduced:

With the better characteristics combination of selecting to algorithm behind the running process, be placed in the L table, j representes that the j of L is capable, at this moment j=1:

Carry out the result of feature selecting after table 4 sequence to algorithm

' 1 ' representes that this characteristic is selected, and ' 0 ' to represent that this characteristic does not have selected.

Initialization: the taboo table is set for empty; Empty table is like table 5; It is 2 that taboo length is set, and the number of times of special pardon criterion regulation is 2, and it is 2 that the greatest iteration step number is set; " 0001 " is separated as the tentative optimum solution Bestsofar and the starting point (current locally optimal solution) of iterative search to what algorithm drew in the back, and the fitness function value that adopts the fisher classifier calculated to go out this initial solution is 0.62;

Table 5 taboo table

The end condition in space 1: judge whether to satisfy the greatest iteration step number,, therefore do not satisfy because do not carry out iteration this moment; It then is the iteration starting point of next time with current locally optimal solution;

Generate candidate's disaggregation according to iteration starting point " 0001 ":

First Candidate Set: i=1, the value of cand (1) is 0, changes cand (1) into 1-0, promptly 1, search since second, running into the 4th is 1, it is become 0 finish; The result of first Candidate Set is " 1000 ";

Second Candidate Set: i=2, the value of cand (2) is 0, changes cand (2) into 1-0, promptly 1, search since the 3rd, running into the 4th is 1, it is become 0 finish; Second Candidate Set gets the result and is " 0100 ";

The 3rd Candidate Set: i=3, the value of cand (3) is 0, changes cand (3) into 1-0, promptly 1, search since the 4th, running into the 4th is 1, it is become 0 finish; Second Candidate Set gets the result and is " 0010 ";

The 4th Candidate Set: i=4, the value of cand (4) is 1, changes cand (4) into 1-1, promptly 0, at this moment, the initial value of p is 4, therefore searches since first again, and running into the 1st is 0, and it is become 1-0, and promptly 1 finishes; The 4th Candidate Set gets the result and is " 1000 ";

Because first Candidate Set and the 4th Candidate Set repeat, only get three Candidate Sets and get final product this moment.Be " 1000 " " 0100 ", " 0010 ";

Optimizing: each fitness function value of separating of calculated candidate disaggregation, " 1000 ", " 0100 "; The fitness function value of " 0010 " is respectively 0.63,0.55,0.65; The fitness function value that can find out " 0010 " is maximum, and " 0010 " as current locally optimal solution, is compared " 0010 " fitness function value with Bestsofar; Find that " 0010 " fitness function value is big, then with " 0010 " as Bestsofar;

Upgrade the taboo table: it is following that " 0010 " is write the taboo table.

Table 6 taboo table

?0	0	1	0

This moment, an iteration was 1 time, did not satisfy end condition, with the starting point of current locally optimal solution as next iteration; Promptly begin to carry out iteration once more, obtain three candidate's disaggregation " 1000 " " 0100 " " 0001 ", carry out after the fitness function value calculates from " 0010 "; Find that " 0010 " fitness function value is maximum; But select a relatively poor candidate solution " 1000 " as locally optimal solution this moment " 0010 " in the taboo table, and itself and Bestsofar are compared; The fitness function value of finding " 1000 " is less, does not replace the value among the Bestsofar;

Upgrade the taboo table, " 1000 " added the taboo table,

Table 7 taboo table

1	0	0	0
				0	0	1	0

Iteration is carried out in " 1000 ", obtain candidate's disaggregation " 0100 " " 0010 " " 0001 ", carry out after the fitness function value calculates; Find that " 0010 " fitness function value is maximum, separating in itself and the taboo table done comparison, find that it is Already in the taboo table; Satisfy the special pardon criterion, discharge " 0010 " in advance, its starting point as locally optimal solution and iteration; Locally optimal solution and Bestsofar are compared, find unanimity, need not to replace Bestsofar; And " 0010 " put into the taboo table once more, at this moment, the taboo table is updated to

Table 8 taboo table

0	0	1	0
				1	0	0	0

Iteration 3 times, meet the end condition of single search volume, be placed on first row of table S to Bestsofar, as shown in the table:

Table 9 is placed the table that preferably separate in each space

Forward the space search of j=2 to: obtain second space separate for

Table 10 is placed the table that preferably separate in each space

Forward the space search of j=3 to, obtain separating of the 3rd space,

Table 11 is placed the table that preferably separate in each space

Forward next space search: j=4 to, j equals the total dimension of characteristic, finishes.

Separating of each space relatively: the fitness function of selecting " 0111 " is maximum, thus with it as optimum solution (Y), output optimum solution " 0111 ".

6.GSR signal emotion recognition effect analysis

Validity feature or characteristics combination are analyzed comparison.The data sample that random division collects obtains training, test according to 3: 1 ratios.Be divided into two parts to every type of emotion physiological reaction sample respectively: randomly draw 3/4 formation X wherein _A*n, be used for training classifier; Constitute T at remaining data _B*n, be used for test, accomplish feature selecting; Wherein a, b are number of samples, and n is an intrinsic dimensionality.And regulation divides time-like that affection data is divided into target emotion and non-target emotion, and when being the target emotion like the data of frightened emotion, non-target emotion is the data of other five types (glad, surprised, detest, sad and indignation) emotions.Sum up characteristic or the characteristics combination that to distinguish six kinds of affective states at last, draw of the mapping of six kinds of emotions to GSR signal physiological characteristic.

Six kinds of final concrete test recognition results of emotion are as shown in table 11, and table 12 has been listed the characteristic title of end product.

Table 11 test and checking recognition result

Table 12 feature selecting result

Hit rate (TPR) and false declaration rate (FPR) representes that they are defined as respectively:

Table 11 adopts improved TABU search to carry out the result that feature selecting obtains, and comprises the test discrimination of discerning six kinds of emotions and the optimal feature subset dimension that obtains.These characteristics can embody the variation of emotion to a certain extent, and the identification of emotion is had bigger contribution degree.Table 12 is through carrying out the characteristic title of every kind of emotion that feature selecting obtains, and when the characteristics combination of experiment proof employing table 12 was distinguished a certain emotion and non-a certain emotion, recognition effect was relatively good, and intrinsic dimensionality is also less.Therefore, a kind of mapping relations of discerning six kinds of affective states and electrodermal response signal characteristic have been found.Can find out that from above experimental result GSR comprises abundant emotion information, it is feasible adopting the GSR signal to be used for emotion recognition research.

Claims (3)

1. one kind is used for the taboo search method that the electrodermal response signal characteristic is selected, and it is characterized in that: said method comprises the steps:

To algorithm, it is capable to form a N-1 after the employing sequence, the bivariate table L of N row, and wherein N is the total dimension of characteristic of selection, and each row is represented a characteristic, and each is gone and is called a space, and wherein n characteristic, 1≤n≤N-1 are selected to have in n space; The value of each element is with " 0 " or " 1 " expression in the table, and " 0 " representative this element when carrying out feature selecting does not have selected, and " 1 " represents this element selected;

Adopt tabu search algorithm to find the solution to the characteristic of choosing in each space, obtain the table S that separates composition in each space;

Select the maximum final feature selecting result of conduct of fitness function in each space.

2. the taboo search method that is used for the selection of electrodermal response signal characteristic as claimed in claim 1 is characterized in that:

Wherein the step of tabu search algorithm is in each space:

S1: initially establish taboo table T=Φ, taboo length is set, the greatest iteration step number is set; With in this space through the value that obtains to algorithm after the sequence as initial solution; And it as the tentative globally optimal solution Bestsofar in this space and the starting point of iterative search; Be current locally optimal solution cand, calculate fitness function

value of initial solution; Wherein ratA and ratB represent the correct recognition rata of target emotion and the correct recognition rata of non-target emotion respectively;

S2: judge whether to satisfy the stopping criterion of greatest iteration step number,, will put into table S to Bestsofar, finish this SPATIAL CALCULATION as satisfying the calculating that then stops this space;

As do not satisfy, then be the iteration starting point of next time with current locally optimal solution cand;

S3: generation N is individual to have candidate's disaggregation that same characteristic features is selected number with this space;

S4: optimizing: calculate the value of the fitness function f of each candidate solution, concentrate from candidate solution and select maximum the separating of fitness function value,

With this separate with the taboo table in separate comparison; If this is separated not in the taboo table, with the value of wherein bigger value as current locally optimal solution and Bestsofar; Then change S5;

If this is separated in the taboo table, and do not satisfy the special pardon criterion, with second largest the separating of fitness function value as locally optimal solution;

If this is separated in the taboo table, and satisfy the special pardon criterion, in the taboo table this separated in advance discharged, and as locally optimal solution, and with the fitness function value comparison of this fitness function value of separating and Bestsofar, as greater than, then separate as Bestsofar with this; Wherein specially pardon criterion and refer to, certain value in the taboo table stipulated number occurs as locally optimal solution in iterative process;

S5: upgrade the taboo table: current locally optimal solution is write the taboo table, change S2.

3. the taboo search method that is used for the selection of electrodermal response signal characteristic as claimed in claim 2 is characterized in that:

The method that generates candidate's disaggregation in the said S3 step is: i position cand (i) value with cand becomes 1-cand (i) respectively, searches since the next bit of i position, runs into the p position that value equals 1-cand (i), its value is become 1-cand (p) back finish; If P=N, then P continues to look into since 1, runs into the p position that value equals 1-cand (i), its value is become 1-cand (p) back finish;

1≤i≤N wherein, 1≤P≤N, N are the total dimension of characteristic.

Images