Evaluating the Accuracy of Machine Learning Algorithms for Classifying Passive Exoskeleton Movements

Exoskeletons are wearable devices that enhance the physical capabilities of their users. Passive exoskeletons, in particular, assist human movement without requiring input from the user, and help prevent muscular over-strain. These devices can be used in a variety of settings, including rehabilitation, military training, and industrial tasks. This thesis proposes a machine learning algorithm for classifying general movements of a passive upper-limb exoskeletons. In particular, it focuses on recognizing three gestures using neural networks: gait, lifting from a standing position, and lifting from the ground. The goal is to develop an AI model that accurately classifies operator actions based on data from inertial sensors attached to the exoskeleton. Additionally, an “anomaly-detection” algorithm has been implemented to capture deviations from normal patterns of movement, which can be indicative of injury or other impairments. The results demonstrate the potential of machine learning approaches for recognizing certain standard gestures and opening the door to future implementations of self-regulating assistance based on movement or operatore effort. Future work could focus on optimizing the algorithm for specific movement tasks and expanding the range of movements that can be classified. This research has the potential to improve our understanding of the effectiveness of passive exoskeletons in various applications by providing real-time data analysis capabilities, which can also be used for further study by the manufacturers’ R&D teams. The present work is the outcome of an internship with Digital Innovation Hub / Confartigianato Vicenza.