Research On Domain Adaptation For Cross-individual Eeg-based Emotion Recognition

Emotion recognition based on EEG signals has always been the preferred method for studying the brain's response to emotional stimuli with its good objectivity and high temporal resolution.However,due to the obvious non-stationary characteristic of EEG emotion features,it is very difficult to predict the corresponding emotional state of EEG signals by using a fixed classification model,which greatly limits the application range of EEG emotion recognition technology.How to improve the generalization ability of EEGbased classification models on samples derived from different individuals(or period)is one of the most challenging frontier directions in the current EEG-based emotion recognition field.The domain adaptation methods can relax the requirement that the training sample and the sample to be tested must satisfy the same probability distribution by implementing knowledge transfer between different domains,and is the preferred method to solve the above problems.In view of this,this article proposes a series of non-stationary EEG-based emotion recognition methods based on domain adaptation through feature adaptation and discriminant model adaptation strategies,focusing on the characteristics of EEG-based emotion recognition technology,and provides several new methods and ideas for solving cross-subject(or period)EEG-based emotion recognition problems.The main research work completed by the dissertation is as follows:First,the problem of under-fitting existing in the shallow network-based domain adaptation method due to the difficulty in mining the complex structure in the probability distribution of EEG data is studied.An off-line EEG feature adaptation method named subspace constraint stacked auto-encoder(SCAE)was proposed.This method uses the domain adaptation technology based on deep neural network to map the original EEG features to the more compact and abstract subspace,so as to reduce the classification bias caused by the mismatch of probability distribution of EEG data.Further,since the deep neural network based on stacked auto-encoder may cause data distribution differences due to a lack of consistency constraints in the parameter optimization process,SCAE minimizes the maximum mean difference(MMD)in the infinite-dimensional reproducing Hilbert space(RKHS)so that the deep neural network remains consistent throughout the training process,avoiding instability problems that may occur during the nonlinear transformation of the original features.By retraining the classification model using the transformed EEG features,the accuracy of the EEG emotion recognition model cross-subject(or period)can be effectively improved.Experimental results show that the proposed method can use both the nonlinear structure of the stacked auto-encoder and the MMD constraint to ensure that the training domain and the test domain are closer together after transforming EEG feature into abstract space,and can effectively improve the accuracy of EEG-based emotion recognition compared with state of art methods.Secondly,since most traditional EEG-based domain adaptation methods have high computational complexity in the learning of kernel functions and the solution of the intrinsic decomposition problem,which always cannot meet the requirements of on-line EEG-based emotion monitoring scenario,a fast feature adaption method called adaptive subspace feature matching(ASFM)is proposed.This method makes full use of the computational complexity advantages of the subspace alignment algorithm(SA)in solving the problem of probability distribution mismatch.According to the Bergmann divergence,the subspace matrixes of the training domain and the test domain are directly adapted under principal component analysis.This makes ASFM achieve a very low computational complexity to adjust the feature transformation matrix,and then adjust the classification model based on changes in the feature space to improve the classification accuracy.Further,since the traditional SA only consider the inter-domain edge distribution adaptation and ignore the problem of conditional distribution differences,the translation scaling transformation strategy is introduced to increase the similarity of the posterior probability in the data domains and thus to solve the conditional probability mismatch problem of the EEG data.The experimental results show that the proposed method can jointly adapt the marginal probability distribution and conditional probability distribution between EEG domains,and has a higher accuracy in the EEG-based emotion recognition than the traditional classification methods.At the same time,ASFM effectively reduces the computational complexity,and achieves the goal of online EEG-based emotion recognition.Finally,aiming at the problem that the assumption of consistency in the distribution of training data for existing EEG-based feature adaptation methods is not strict,and the possible multi-source problems of training samples are neglected,a multi-subject domain adaptation framework(MSDA)which is the combination of feature adaptation and discriminative model adaptation is proposed.On account of the issue that measuring the discrepancy of data distribution between individuals in the original feature space is inaccuracy,the feature adaptation method is used to obtain the abstract public features in the implicit feature space to quantify the degree of data distribution between individuals.Further,taking into account the potential correlation between training data individuals,a global weight estimation method based on the smooth hypothesis theory is proposed to estimate the importance of the training individual sample set relative to the test individual.On this basis,by adding regular constraints to the objective function to narrow the distance between the target classification parameters and the linear combination of N independent source classifiers,the parameters of the new classifiers can be obtained by solving the convex optimization problem.Experiments show that the proposed method can autonomously adjust the knowledge transfer process according to the similarity between multiple individuals in the training set and the target individual to be tested,and then improve the accuracy of EEG emotion recognition for the target individual to be tested.