A Study Of Affective States Recognition Based On Dual-Structure Particle Swarm Optimization And K-Nearest Neighbors From Physiological Signals

Contemporary science and technology information allows the rapid development of human dependence on the computer constantly strengthens the capacity of human-computer interaction researchers more attention.How to achieve the anthropomorphic computer so that it can be perceived environment,the atmosphere,the attitude and emotion from human,etc.,and to respond timely and in harmony with users when communicating emotionally with them,has become a "human-computer emotional interaction," a field of study hot topics.Let the computer with human emotion,is to first of all,the computer can understand people's emotional states and to improve the computer's ability to identify emotional states,that is,for affective computing. Emotion Recognition is a key issue of Affective Computing and one of the foundations to create a harmonious environment for human-machine.Affective Computing is one of the emerging issues,it is about the emotional aspects of the terms,the purpose is to give the computer the ability to identify,understand,express and adapt to human emotions.Affective Computing emotion recognition is an important component of the study of emotion recognition including facial expressions,voice,gesture,text and physiological signals recognition and so on.Physiological signals are biological signals which accompanied by emotional changes produced by the human internal organs,more objectively reflect the real states of emotion.Through a series of experiments,Ekman showed that at least for some emotions,its physiological responses are specific.Picard of MIT who led the media to prove the application of physiological signals to identify emotion states is feasible and therefore the physiological signals for emotion recognition of further study of the theory provides a reliable support.During emotion recognition,a large number of irrelevant or redundant features tend to affect the recognition speed and accuracy,so we need for feature selection.Feature selection problem is essentially a combinatorial optimization problem,although some scholars put forward a number of search algorithms,but there have none algorithm is regarded as an effective search algorithm so far.Dual-Structure Particle Swarm Optimization(DSPSO) algorithm is a smart group based on the global optimization algorithm,its unique structure makes for consecutive PSO can be used to solve combinatorial optimization problems,and has a simple code,the number of small individual calculate the speed and population diversity,and easy to understand, easy to implement and so on.Therefore,the thesis research applied DSPSO algorithm to feature selection in order to increase the recognition rate of emotion recognition from physiological signals.In the feature search process needed to evaluate the combination of selected feature subsets, the evaluation process is a classification process.As the K-Nerest Neighbors(KNN) classification algorithm fast and efficient,the need to repeatedly call the classification in the feature search process,KNN has its unique advantages.Therefore,the thesis research applied KNN classification algorithm to identify the feature subsets of the search process to enhance the computing speed and effect.In this paper,we use the data of Multimedia and Signal Processing laboratory,Augsburg University,Germany,as an example first(four physiological signals(Electromyogram(EMG), Skin Conductivity(SC),Electrocardiogram(ECG),Respiration(RSP)) based on four emotions (Joy,Anger,Sadness,Pleasure)).The physiological signal-based emotion recognition was investigated.We used the data to do the following four areas:(1) Aim at feature redundancy problem of emotion recognition from physiological signals, research will be to introduce the idea of computational intelligence to the feature selection of the physiological signals with a view to prove that it can improve the correct recognition rate. DSPSO-KNN was apllied to select features,single physiological signals and a variety of physiological signals are used to identify signle emotions and a variety of emotions. (2) For the bad effect when DSPSO-KNN processes major number of original features,the thesis used chaos variation,niche and multi-population DSPSO-KNN to select features that aim to improve the effect of feature selection when the original set is too large.We studied the single physiological signals to identify the single emotions and a variety of emotions,at the same time, we used avariety of physiological signals to identify the single emotions and a variety of emotions.(3) This paper proposed incremental K for avoiding indivisibility about multi-classification. In view of repeated emergence about same swarms when iteration tends to be convergent, look-up table method is presented to avoid superfluous calculation.(4) This paper compared the effect of emotion recognition between DSPSO-KNN and traditional Sequential Floating Forward Selection(SFFS) and Sequential Floating Backword Selection(SFBS) methods in order to incarnate the advantages of DSPSO-KNN.In order to further research work,we tested 391 college students(Electrocardiogram(ECG) based on two emotions(Joy,Sadness)),and use 150 samples better to do the following work:After preprocessing the ECG data,we used wavelet transform to locate P-QRS-T wave and extracted statistical features,then used these features to do emotion recognition;Thesis confirmed the correctness of the above-mentioned work through a large number of simulation experiments,and achieved the following research areas:(1) 193 features were extracted from four physiological signals,and DSPSO-KNN was used to select features.Within the single physiological signals,the EMG signal got the best effect to identify the four emotions which reached to 83%;SC signal got the worst effect,only 51%;and the recognition rate reached 93%when applied four physiological signals to identify four states of emotion.(2) Learned through the simulation experiment,we got the fallowing points:improved DSPSO-KNN has no advantage when the original feature set is small,but it has better effect when the original set is large.For instance,EMG and SC both have 21 features,so the effect of emotion recognition had not improved;ECG has 84 features.The recognition rate was 69% before improvement and 73%after improvement.We find out that some features indeed have special contribution when identifying some emotions.The median value after second-order difference of EMG signal has been always selected when just used EMG signal to identify pleasure state or used all these four signals to identify pleasure state.(3) Incremental K of KNN algorithm for multiclass classification not only solved the problem that these classes can not be separated,and the recognition rate has been enhanced,for example,to identify four types from EMG the emotion recognition rate reached 83%,while traditional methods were 79.4%and 80%;Look-up table method has cut down compute time from 75.54 seconds to 15.21 seconds when use EMG signal to identify four emotion states and fixed iteration with 200.(4) Identification four emotional states by four physiological signals,SFFS got the best recognition rate of 93%,33 features,SFBS got the best recognition rate of 92%,23 features, improved DSPSO-KNN got the best recognition rate of 93%,12 features.(5) We used the ECG signal gained through the laboratory to identify happy and sad feelings.The optimal subset includes 7 features,and the recognition rate reached 87.33%.Through the study of this article,we know that the combination of DSPSO and KNN (DSPSO-KNN) methods is a effective feature selection method,through the use of the method can effectively improve the rate of emotional recognition and to reduce the calculation.This paper lays the theoretical foundation for the application of emotion recognition.