Enhancing Human-Centric Physiological Data-Driven Heat Stress Assessment in Construction through a Transfer Learning-Based Approach

Recent advances in physiological sensors and machine learning have led to the development of non-invasive heat stress monitoring frameworks that can continuously and objectively assess the heat stress levels of workers in the field by analyzing their physiological data. However, variations in the statistical distribution of physiological data due to individual differences in responses to stressors negatively impact the accuracy of the assessment. To address this issue, this study proposed a transfer learning-based framework to improve the performance of non-invasive heat stress monitoring. The framework utilizes autoencoder and domain adaptation-based transfer learning techniques to reduce the deviation of the statistical distributions of physiological data across different individuals, leading to a more robust assessment of workers' heat stress levels. To evaluate the effectiveness of the framework, physiological data was collected from 14 subjects performing roofing tasks with different heat stress exposure levels (low, medium, and high). Results showed that the proposed framework had a more robust performance on physiological data with distributional shifts, achieving an accuracy of over 89.9% in assessing heat stress levels across different subjects, a 6.3% improvement compared to existing frameworks. This study contributes to the advancement of heat stress assessment for construction workers.