Purpose Falls from heights (FFH) remain the leading cause of fatalities in the construction industry worldwide. Mental fatigue associated with heat stress and psychological stressors is a critical yet overlooked contributor to FFH risk. Current risk-mitigation approaches emphasize physical protective measures but fail to address cognitive impairment caused by environmental and psychological stressors. Moreover, existing deep-learning (DL) models struggle to capture the oscillatory patterns and long-term dependencies inherent in time-series physiological data. This study develops a novel prediction model evaluated on two case studies: (1) FFH risk associated with heat stress and hazardous areas and (2) mental fatigue classification in construction workers. Design/methodology/approach This study proposes the Bidirectional Revolving Gate Fourier Transform (BiRGFT), a novel DL architecture that integrates bidirectional gated recurrent units with dual bidirectional fast Fourier transform and cross-attention mechanisms. The model leverages data from wearable sensors, environmental monitoring systems and historical fall incident records at active construction sites. Findings BiRGFT achieved average precision, recall and F1-score of approximately 0.98 for case study 1 and 0.95 for case study 2, outperforming eight baseline DL and hybrid DL models. The minimal gap between training and testing results indicates strong generalization. The results support the capability of this real-time risk classification system to enable continuous monitoring and generate timely alerts for safety managers. Originality/value This research advances DL applications in construction safety by bridging time-frequency domain representations through BiRGFT. This framework enables the integrated monitoring of environmental risks and worker mental states, supporting proactive safety management across diverse construction environments.
Cheng et al. (Mon,) studied this question.