Cross-subject recognition of operator functional states via EEG and switching deep belief networks with adaptive weights

Assessing operator functional states (OFS) by using neurophysiological signals can provide continuous prediction of instantaneous human performance in safety-critical human-machine systems. Most existing OFS recognizers were built via a subject-dependent manner, where a new model has to be trained based on historical physiological data of the same subject. The main objective of this paper is to generalize such paradigm to cross-subject OFS recognition by exploiting the new improvements in deep learning principles. To this end, we propose a novel EEG-based OFS classifier, switching deep belief networks with adaptive weights (SDBN), which is generic for detecting variations of mental workload, mental fatigue, and the coupling effect of the two variables across multiple subjects. The temporal OFS is predicted by switching the ensembles of the static and adaptive DBNs at each time step via a Gaussian-kernel based criterion. The results comparison demonstrates that the SDBN not only significantly improves classification accuracy but also has the capability to distinguish multiple dimensions in OFS.