文献类型：会议论文

英文题名：Recognition of Cognitive Task Load Levels Using Single Channel EEG and Stacked Denoising Autoencoder

作者：Yin Zhong[1];Zhang Jianhua[2]

机构：[1]Univ Shanghai Sci & Technol, Shanghai Key Lab Modern Opt Syst, Minist Educ, Engn Res Ctr Opt Instrument & Syst, Shanghai 200093, Peoples R China;[2]East China Univ Sci & Technol, Dept Automat, Shanghai 200237, Peoples R China

会议论文集：35th Chinese Control Conference (CCC)

会议日期：JUL 27-29, 2016

会议地点：Chengdu, PEOPLES R CHINA

语种：英文

外文关键词：Cognitive task load; mental workload; electroencephalogram; deep learning; operator functional state

摘要：Evaluation of operator Cognitive Task Load (CTL) level is quite crucial in Human-Machine (HM) collaborative task environment since operator mental overload or inattention caused by abnormal CTL states may lead to human performance degradation or even catastrophic accidents. One of the most practical approaches tackling this issue is to use ongoing electroencephalogram (EEG) in which human cognitive state can he objectively estimated. However, the accurate recognition of CTL via single channel EEG with the lowest-intrusivity to task condition is particularly challenging as EEG is characterized by individual dependency and nonstationarity. In this paper, a deep learning model designed by Stacked Denoising AutoEncoder (SDAE) is employed on single EEG channel signal to estimate binary levels (low vs. high) of CTL. By adopting a simulated HM process control system, the operator EEG data for 8 healthy subjects under different task demands were collected on two experimental sessions across two consecutive days. Based on the computed full power spectral of EEG, the number of nodes in SADE is determined by greedy search according to the optimal training error of each layer. The shallow layers of the designed deep network are used to extract the subject-specific information related to CTL variation while the stable power features were reconstructed in those deep layers. Finally, the proposed method is demonstrated to be effective and 74% classification rate across sessions in average of all subjects were achieved.