Sleep Analysis for Fatigue of Crewmembers with Deep Multi-Task Learning

This paper targets to investigate the potential performance of crewmembers given the sleep quality surveillance. The fatigue of crewmembers can significantly affect flight safety and service quality. The quantification and prediction of fatigue is highly desired in practice, and lead to better control of the crewmember arrangement and scheduling. Though the working status can be related to many inherent causal factors, sleep quality plays a major role in fatigue. In this work, we propose to develop an end-to-end system to collect the sleep data with wearable devices, and explore its relations with the popular fatigue indexes, e.g., Karolinska Sleepiness Scale (KSS), Psychomotor Vigilance Task (PVT), and NASA Task Load Index (NASA-TLX). Specifically, we collected the sleep time and quality measurement of 256 crewmembers for three days before the flight, and the corresponding KSS, PVT, and NASA-TLX during or just after the flight. Then, a multi-task deep learning framework is developed to learn the mapping from sleep to fatigue indexes. We trained and evaluated on 200 and 56 crewmembers with the quantified results showing that fatigue is highly predictive with the sleep data. The learned deep learning model can potentially be applied to the pre-flight crewmember screen to avoid the fatigue driving and service.