Human Environment Interaction Featuring Activity Recognition in Workers’ Wearable Technologies and Industrial Exoskeletons

Industry 5.0 integrates the concept of human centricity into the digital industrial paradigm, thereby emphasizing three core values: resilience, sustainability, and human-centricity. This approach can be particularly valuable in the industrial sector, which plays a vital role in economic growth but faces ongoing challenges, including health risks, financial losses, and low productivity. Industrial environments require continuous physical effort, often leading to health risks, such as work-related musculoskeletal disorders (WMSDs). With the growing integration of wearable technologies and robotics, Human Activity Recognition (HAR) has emerged as a critical tool for creating safer, adaptive, and human-centred workplaces. This PhD research aims to develop HAR algorithms for deployment on portable devices compatible with edge computing. The ultimate goal is to integrate HAR into smart clothing and occupational exoskeletons (OEs) to provide applications in the field of Human-Environment Interaction (HEI) and Human-Robot Interaction (HRI) in industrial settings, such as fatigue estimation, ergonomic performance evaluation, and HAR-driven OEs control, enabling multi-task assistance. To develop an efficient HAR algorithm for portable devices, key factors such as dataset design, accuracy, real-time performance, and computational efficiency were examined. An elementary-transition-elementary (ETE) approach was introduced to segment complex activities into elementary movements separated by transitions, reducing labelling errors and enabling flexible dataset expansion. Sensor configuration, signal selection, and feature extraction were optimized to improve accuracy, while the window size was adjusted to balance inference time and real-time suitability. Finally, machine learning (ML) and deep learning (DL) models were compared for computational feasibility on edge devices, and the developed models were applied to worker performance evaluation, fatigue estimation, and OEs control strategies. The Elementary-Transition-Elementary (ETE) approach effectively reduced human labelling errors, simplifying data annotation and dataset construction. Various sensor configurations were explored without hardware constraints, using multi-modal data, i.e., accelerations and joint angles, to enhance model robustness. A window size of 0.4s with 50% overlap was selected, enabling 10 real-time predictions per second. Among ML and DL models, the neural network proved most lightweight at only 45 kB. The developed models were applied across several domains: in worker performance evaluation, they distinguished safe (squat) from risky (stoop) lifting techniques; in fatigue assessment, the ML-based regression model achieved a 7.34% MAE, outperforming prior studies, while the combined HAR–physiological model achieved < 2% MAE. For OEs control, HAR-based models exceeded 90% accuracy across manual material handling tasks, demonstrating strong potential for assistive control applications. This PhD research demonstrates the potential of HAR applied to wearable technology, enabling human-centred, intelligent, and adaptive systems for industrial environments. By linking HAR with physiological models, we can provide fatigue estimation in HEI to improve worker well-being. Additionally, by applying HAR in robotic assistance, we develop advanced adaptive HRI in OEs multi-task assistance. These applications align with the principles of Industry 5.0 to create a safer, more efficient, and human-friendly workplace.