Cross-session classification of mental workload levels using EEG and an adaptive deep learning model

Abstract

Abstract Evaluation of operator Mental Workload (MW) levels via ongoing electroencephalogram (EEG) is quite promising in Human-Machine (HM) collaborative task environment to alarm the temporal operator performance degradation. However, accurate recognition of MW states via a static pattern classifier with training and testing EEG signals recoded on separate days is particularly challenging as EEG features are differently distributed across different sessions. Motivated by the superiority of the deep learning approaches for stable feature abstractions in higher levels, an adaptive Stacked Denoising AutoEncoder (SDAE) is developed to tackling such cross-session MW classification task in which the weights of the shallow hidden neurons could be adaptively updated during the testing procedure. The generalization capability of the adaptive SDAE is first evaluated under within/cross-session conditions. Then, we compare it with the state of the art MW classifiers under different feature selection and the noise corruption paradigms. The results indicate a higher performance of the adaptive SDAE in dealing with the cross-session EEG features. By analyzing the optimal step length, the data augmentation scheme and the computational cost for iterative tuning, the adaptive SDAE is also demonstrated to be acceptable for online implementation.