Optimizing CNN Hyperparameters for Mental Fatigue Assessment in Demanding Maritime Operations

Abstract

Human-related issues play an important role in accidents and causalities in demanding maritime operations. The industry lacks an approach capable of preventively assessing maritime operators' mental fatigue and awareness levels before accidents happen. Aiming to reduce intrusiveness, we focused on improving the mental fatigue assessment capabilities of a combination of electroencephalogram and electrocardiogram sensors by investigating the optimization of convolutional neural networks by Bayesian optimization with Gaussian process. We proposed a mapping function to optimize the network structure without the need for a tree-like structure to define the domain of variables for the optimization process. We applied the proposed approach in a simulated vessel piloting task. Even though the mental fatigue assessment for the cross-subject case is a complex classification task, the trained convolutional neural network could achieve good generalization performance (97.6% test accuracy). Finally, we also proposed a method to improve the depiction of the mental fatigue build up process. The framework presented in this work can contribute for reducing accident risk in maritime operations by improving the accuracy and assessment quality of neural network-based mental fatigue assessment tools.