Passive BCI Towards Health and Safety in Industry: Forecasting Human Vigilance 5.5 s Ahead

Brain-computer interfaces based on electroencephalography (EEG) recordings are gaining increasing interest in the industrial domain, aiming to enhance health, safety and performance by optimizing the cognitive load of industrial operators and facilitating human-robot interactions. This study introduces a novel experimental protocol and analysis pipeline for predicting vigilance degradation during repetitive tasks. A dataset was recorded from 10 volunteers who observed a robotic arm executing three distinct movements. The EEG power spectrum was analyzed over time using the continuous wavelet transform. Upon verifying the increased amplitude of EEG oscillations in the 8-12 Hz frequency band, we forecast its behaviour, comparing the vector autoregressive model with two deep learning recurrent architectures. The proposed encoder-decoder gated recurrent unit model obtained accurate forecasts (mean absolute error = 0.048, R² = 0.726) up to 5.5 s into the future. The findings suggested the feasibility of vigilance monitoring in the Industry 5.0 framework, proposing a strategy to prevent human accidents and performance decline during monotonous activities.