Physiological correlates and multimodal recognition of hazard-perception performance in train drivers under abnormal...

Human error remains a persistent concern in railway safety, and whether a train driver notices a track hazard in time depends on the driver's functional state, which physiological signals can track. Studies of train drivers have typically examined one state at a time, leaving it unclear whether different states impair the same aspect of hazard perception. Physiological change is seldom related to the type of failure that follows. This study induced emotion, distraction and fatigue in the same 21 train drivers operating a full-scale CRH380B simulator, with scalp EEG, photoplethysmography and electrodermal activity recorded throughout. Every trial was scored as correct, miss or mislocalisation, and six physiological indicators and performance were modelled jointly in Bayesian multilevel models at the trial and block levels. Miss rates rose under emotion and distraction and at the highest fatigue level, while mislocalisation was largely unaffected. A selectivity contrast from the same posterior showed that under distraction the impairment fell on miss rather than on mislocalisation. Distraction produced the strongest impairment, with miss odds nearly four times baseline and the largest increases for tasks requiring effortful internal computation. Under fatigue physiological changes were detectable at earlier fatigue levels than behavioural deterioration. Peripheral measures tracked state changes the P3 did not register, whereas the P3 was most closely associated with failure type. The miss rate and the response speed were combined into a single failure rate, on which the driver states were stratified into three performance levels. A gated spatial-filter and hierarchical-attention fusion network (GSHA-Net) combining single-trial EEG with peripheral features recognised these levels within a session with a macro F1 of 0.96, outperforming the baseline models.