Automatic drowsiness detection for safety-critical operations using ensemble models and EEG signals

Abstract

Listen

Translate article

Recently, industrial sectors that stage occupational and environment safety critical tasks, such as the oil and gas industry, have been interested in monitoring biological parameters to prevent human errors and enhance process safety with emergency preparedness and response. In this context, human reliability plays a fundamental role to avoid possible catastrophic accidents triggered by human factor, for example workers’ fatigue. Drowsiness, as a main causes of fatigue, maybe recognized through patterns in electroencephalogram (EEG) signal. In this paper, we propose a drowsiness recognition system that combines information from different EEG signal channels and machine learning in an ensemble methodology, novel for this context. We consider two ensemble approaches: the bagging, using five and three channels, and the voting, using a single channel. To validate the proposed system, DROZY, a real and public database containing drowsiness data, was used in three cases: (1) evaluated in all available subjects; (2) evaluated in specific subjects with general model; and (3) evaluated for specific subjects and dedicated models. The results show that our proposed system has high accuracy above 90%, in most subjects for Case 3. While for Cases 1 and 2, the ensemble model is equivalent to the best results of the classifiers from the single-channels. Furthermore, collecting many channels of EEG signals is often expensive and cumbersome for humans, and the schemes using many channels of EEG signals do not necessarily lead to better performances.