Worker-centric heat strain analysis: Integrating physiological signals with ensemble learning and domain adaptation

Abstract

Despite advances in utilizing physiological sensors and machine learning (ML) algorithms, accurately and consistently monitoring heat strain levels of field workers on job sites remains a challenge not fully addressed by previous research. Existing frameworks often fail to adapt to the diverse physiology of workers and their varied working conditions, leading to concerns about stability, reliability, and accuracy. To address these limitations, a worker-centered heat strain monitoring framework was introduced, leveraging physiological data from wearable biosensors for a more accurate and consistent estimation of heat strain risks. A high-fidelity virtual reality (VR) environment was developed to simulate heat-vulnerable occupations for quality data collection. Building on this foundation, EnsmTrBoost, a physiological sensing framework integrating ensemble learning and domain adaptation, was developed. This framework exhibited a remarkable prediction accuracy exceeding 93%. This paper advances heat strain monitoring and supports the development of early warning systems for heat-related fatalities at work sites.