Machine learning aided smart helmet: wearable sensors for workplace safety and health (WSH)​

Abstract

To ensure the safety of those in the construction sector, one of Singapore’s strategies is the implementation of technology-enabled solutions as outlined in the Workplace Safety and Health (WSH) 2028 Report. A subset of these solutions focuses on the use of smart personal protective equipment (PPE), which aims to overcome any potential misuse and disuse of PPE in the workplace. A number of studies have examined this problem and proposed various solutions, including surveillance cameras which were limited by blind spots, and mounted cameras which faced increased costs and ergonomic problems. A previous study observed that using micro-electromechanical systems (MEMS) based sensors to capture data and parsing said data through multiple machine learning (ML) models is a feasible solution. These ML models included logistic regression (LR), decision tree (DT), support vector machine (SVM) and Gaussian Naïve Bayes (GNB). In this study, the aim is to investigate the prolonged thermophysical response of ambient-microclimate humidity difference rate of change (AMHDROC) in a controlled environment and its relevance to the helmet monitoring algorithm. Other important parameters were also analysed: (a) internal humidity; (b) external humidity; (c) ambient-microclimate humidity difference (AMHD), which is the sole predictor feature used in the ML models. Additionally, 3 more classification ML models were added to evaluate their usefulness in the helmet monitoring algorithm, namely K-nearest neighbours (KNN), gradient boosting (GB) and random forest (RF). The data shows that the SVM model performs the most consistently with an accuracy mean of 92.94%, a precision mean of 89.25%, an F1-score mean of 93.94% and mean lag time of approximately 44 ticks. The least robust model was determined to be the GNB model which recorded an accuracy mean of 54.35%, a precision mean of 53.36%, an F1-score mean of 70.43% and mean lag time of 289 to 290 ticks. This shows that in order of relative performance from most to least satisfactory, SVM is closely followed by LR, GB and RF in one group, then DT and KNN in another, ending with GNB. These findings show that AMHD alone is a reliable feature for prediction, and the simultaneous use of multiple ML models can be implemented with MEMS-based sensors to accurately predict the wearing of PPE, which opens the possibilities to developing similar low-cost smart PPE for construction workers.