Study On Cognitive Computing Of Auditory Attention Decoding And Emotion Recognition Based On Electroencephalography

With the development of intelligent science,artificial intelligence has made rapid progress.In some application,artificial intelligence has demonstrated the ability to solve problems beyond human beings.In recent years,the use of cognitive computing methods to develop brain-like artificial intelligence technology has become the focus of researchers.The existing artificial intelligence technology,with the help of big data mining and machine learning methods as a representative of deep learning,has achieved a certain degree of brain-like intelligence and emotional intelligence.However,it still has not achieved leapfrog technological development.Faced with tasks such as auditory recognition,speech understanding,and emotional decision-making in complex acoustic scenes,the existing signal processing methods and technologies are still powerless.However,such tasks can be often done easily for human beings who rely on their intelligence such as psychological awareness,attention,emotional cognition,etc.This kind of human brain intelligence is beyond the reach of existing artificial intelligence technology.In this thesis,taking auditory attention and psychological emotions as the research object,the main research purpose is to explore the mechanism of cognitive computing in mind for mental states decoding from EEG signals.The thesis is to develop the method and technical realization of cognitive computing on auditory object-specific attention identification,auditory selective attention decoding and cross-individual emotion recognition,which can give the cognitive computing system with some brain-like intelligence like human beings to predict the relevant psychological states.The main research content of this thesis includes the following points:Firstly,we give a deep study on the decoding approaches of EEG signals.In view of the technical advantage that the combination of time series analysis methods of decomposition and reconstruction and entropy measures can effectively improve the performance of EEG decoding,a new decoding approach of EEG signals is proposed,which is based on singular spectrum analysis(SSA)and entropy measures.First,the decoding method uses the SSA method to decompose the EEG signals into SSA components;second,based on the decomposed SSA components,using entropy measures to extract the EEG entropy features;finally,using support vector machines as the pattern classifier to perform the EEG decoding task.The experiments use EEG signals of different eye states as an example to evaluate the performance of the proposed EEG decoding approach.The research results demonstrated that the proposed EEG decoding approach can effectively improves the accuracy of EEG decoding and achieves the performance optimization of EEG decoding tasks.Secondly,we conduct a comprehensive study on the fast computation method of entropy measures.In order to accelerate the computation of entropy measures such as approximate entropy(Ap En),sample entropy(Samp En),multi-scale entropy(MSE),aiming at enhancing theirs application potential in the study of EEG decoding,an effective method to accelerate the computation of these entropy measures is proposed by exploiting vectors with dissimilarity(VDS)in the thesis.This proposed fast conputatuon method optimizes the most time-consuming step of vectors distance calculation in entropy measures.Before the calculation of vectors distance,a vector dissimilarity judgment criterion is constructed to judge the dissimilar vectors.The experiments use the simulated signals and real EEG data to perform the calculation of Samp En,Ap En,time-shift multi-scale entropy(TSME),ao as to evaluate the performance of the proposed fast computation method.Experiment results show that when compared with the conventional method,the prposed fast calculation method can effectively accelerate the computation of Samp En,TSME and Ap En.Thirdly,we conduct an in-deep research and analysis on the auditory attention decoding based on single-trial EEG signals.First,based on the combination of EEG entropy measures and machine learning approaches,an identification method of auditory object-specific attention from single-trial EEG signals is proposed.In the experimental study,an auditory experiment with three kinds of auditory object-specific attention is designed,and the corresponding EEG data are collected from 13 subjects.Experiment results show that the proposed method can effectively achieve the identification task of auditory object-specific attention from single-trial EEG signals.On this basis,the thesis proposes a new EEG decoding method of auditory selective attention based on LSTM model.In the experimental study,an auditory experiment of two-speaker dichotic listening paradigm is designed,and the corresponding EEG data are collected from 21 subjects.The experiment results showed that the targeted speech of auditory selective attention based on LSTM RNN was identified with the best accuracy by the proposed decoding method of auditory selective attention based on LSTM RNN,and achieved high-accuracy EEG decoding of people's auditory attention state from EEG signals.Finally,we conduct an in-deep research and analysis on the cross-individual psychological emotional states recognition from EEG signals.Due to the individual differences of EEG emotion responses,most existing emotion recognition methods based on EEG signals have weak universality and poor generalization ability.In order to overcome these shortcomings,a new dynamic sample entropy-based pattern learning to identify emotions from EEG signals across individual is proposed in this thesis.The experiments used the EEG emotion dataset SEED to carry out cross-individual emotion recognition for 15 subjects.By comparing with the existing related studies,the experiment results showed that the proposed dynamic sample entropy pattern learning achieves better cross-individual emotion recognition,and has better universality and generalization ability.The cognitive computing method for cross-individual emotion recognition developed in this thesis has realized the optimization and innovation of EEG emotion pattern recognition,and can be used to reasonably predict people's psychological emotion states from EEG signals.In this thesis,the cognitive computing study for auditory attention decoding and cross-individual psychological emotions recognition from EEG signals can give the artificial intelligence system to have some cognitive intelligence with attention mechanism and emotion perception intelligence,which can be used to predict the related mental states.