A feedforward neural network for drone accident prediction from physiological signals
Academic Article
Overview
Research
Identity
Additional Document Info
Other
View All
Overview
abstract
PurposeAs drones are rapidly transforming tasks such as mapping and surveying, safety inspection and progress monitoring, human operators continue to play a critical role in ensuring safe drone missions in compliance with safety regulations and standard operating procedures. Research shows that operator's stress and fatigue are leading causes of drone accidents. Building upon the authors past work, this study presents a systematic approach to predicting impending drone accidents using data that capture the drone operator's physiological state preceding the accident.Design/methodology/approachThe authors collect physiological data from 25 participants in real-world and virtual reality flight experiments to design a feedforward neural network (FNN) with back propagation. Four time series signals, namely electrodermal activity (EDA), skin temperature (ST), electrocardiogram (ECG) and heart rate (HR), are selected, filtered for noise and used to extract 92 time- and frequency-domain features. The FNN is trained with data from a window of length t=38 s to predict accidents in the next p=38 s.FindingsAnalysis of model performance in all 36 combinations of analysis window (t) and prediction horizon (p) combinations reveals that the FNN trained with 8 s of physiological signal (i.e. t=8) to predict drone accidents in the next 6 s (i.e. p=6) achieved the highest F1-score of 0.81 and AP of 0.71 after feature selection and data balancing.Originality/valueThe safety and integrity of collaborative humanmachine systems (e.g. remotely operated drones) rely on not only the attributes of the human operator or the machinery but also how one perceives the other and adopts to the evolving nature of the operational environment. This study is a first systematic attempt at objective prediction of potential drone accident events from operator's physiological data in (near-) real time. Findings will lay the foundation for creating automated intervention systems for drone operations, ultimately leading to safer jobsites.
