Human Activity Recognition and Payload Classification for Low-Back Exoskeletons Using Deep Residual Network

Abstract

Exoskeletons created by robotics currently hold significant potential for improving the well-being and quality of life for manufacturing laborers. However, the effectiveness of existing solutions is often hindered by inadequate control systems. The limited adoption of exoskeletons in occupational settings can be attributed to their inability to adapt to users and tasks, which is a result of the constantly changing circumstances in which they are utilized. To address this issue, wearable sensors capable of being carried on the human body can employ learning-based methodologies to anticipate and classify user motion characteristics by training on extensive datasets. This study introduces a deep learning approach that utilizes inertial sensors to equip manufacturing exoskeletons with the ability to recognize human movements and adjust for payload compensation. Inertial measurement equipment can be conveniently integrated into manufacturing exoskeletons through either wearable or embeddable means. A deep residual network was employed for human action identification and categorization of lifted object weights up to 15 kg. The study achieved an average F1-score of 93.91% for activity identification and 94.35% for payload classification using the provided model. The model was trained and tested on a sample of 12 young, healthy volunteers.