Research On Mental Workload Classification Method Based On Subject-Specific Model

Human-computer interaction has gained significance due to the swift advancement of science and technology.In the development and evaluation of intricate man-machine systems,mental workload(MW)plays a crucial role as it affects operator performance.Accurate identification of MW is of great significance to the safety of operators and the efficient operation of the system.The brain is an organ responsible for information processing and decision-making.MW level has a direct impact on the efficiency of brain work and can be measured through electroencephalogram(EEG)activity.Analysis based on EEG signal characteristics is a common method for MW classification.At present,most of the research on MW classification uses machine learning methods.However,EEG signals are random and non-stationary.There may be some differences in EEG signals collected from different subjects and different environments.Modeling the EEG signals in the subjects directly will reduce the generalization ability of classifier,and it is difficult for the model to correctly classify the new subject data.In response to the above problems,this study records the EEG signals of the labeled subjects as the source domain,and records the EEG signals of the unlabeled subjects as the target domain.This thesis focuses on the differences in quantity between the source domain and the target domain,and the main content is as follows:(1)In the scenario where there is only one source domain and one target domain,this thesis compares the accuracy of MW identification using three different domain adaptation methods,among which the subject-specific MW classification model that utilizes correlation alignment achieved the better result..This method aligns the highorder statistics of different subjects’ EEG signal features through a linear mapping from source domain to target domain,which aims to minimize the differences in the distribution of EEG signals across different subjects.This method enhances the model’s generalization ability and classification accuracy for MW in comparison to conventional machine learning techniques.(2)For the problem of multi-source domains,this thesis proposes a subjectspecific MW classification based on multi-source domain adaptation(MSDA).This method uses a parallel alignment strategy for multiple source domains with the target domain in order to augment the available training data,and at the same time use the source domain selection algorithm based on Kullback-Leibler(KL)divergence to calculate the distribution difference between the two domains,this method ranks and selects source domains by setting a threshold,and uses the source domain closest to the target domain as a training sample to construct a subject-specific MW classification model.The accuracy of MW classification is further enhanced compared to the single source domain model.(3)To enhance the classification of subject-specific MW across multiple sources,this thesis constructed a Multi-Source Domain Adaptation Network(MSDANet).This model using deep domain adaptation methods.The time-domain EEG signals were transformed into 2D images as inputs to the network,and a convolutional neural network with fewer parameters was used as the model framework to avoid overfitting.Additionally,this article added the objective function of correlation alignment methods and KL divergence as loss functions to the network.This thesis also discussed the changes in classification accuracy of the MW classification model as the number of source domains increased,clarifying the usage scenarios for MSDA and MSDANet.Finally,the average classification accuracy of the personalized mental workload classification model based on deep domain adaptation reached 93.47%.