Captured EDA, HRV, and PPG of exoskeleton users using a multimodal biosensor (Emotibit), pre-processed the signals, and extracted features to predict physical fatigue using machine learning.
Abstract
Despite the benefits of active back-support exoskeletons, their use may lead to unintended consequences that could exacerbate physical fatigue among users, potentially compromising safety, especially in rugged environments like construction sites. Recognizing instances and levels of physical fatigue experienced by construction workers while using active back-support exoskeletons is crucial for developing strategies to mitigate associated risks. This study investigated supervised machine learning models for classifying physical fatigue using physiological data such as electrodermal activity (EDA), photoplethysmogram (PPG), and skin temperature (ST), as well as psychological brain data from electroencephalogram (EEG) with carpentry framing task as a case study. The top-performing classifiers achieved accuracy of 88.9%, 85.5%, 90.3%, and 89.3% for EDA, PPG, ST, and EEG data, respectively, using techniques such as Ensemble and Support Vector Machine. Synthetic Minority Over-sampling Technique data augmentation algorithm was adopted to improve the accuracy of the models, resulting in increases of 10.9%, 7.7%, and 10.7% in classification accuracy for EDA, ST, and EEG data, respectively. Additionally, key features extracted from each data type are highlighted. This study contributes to the scarce literature on predicting physical fatigue levels among exoskeleton users, potentially improving productivity by strategically reallocating fatigued workers. The developed models can potentially enhance the safety use of exoskeletons, aiding manufacturers in device assessment. The variation in performance across different data types highlights the importance of integrating multiple data sources, particularly EEG and ST, to improve the accuracy and reliability of fatigue monitoring systems. The identified top-ranked features could guide future studies exploring machine learning models for classifying physical fatigue.
Methodology
