A Machine Learning-Based Approach for Lifting Load Estimation

Abstract

Lifting weight estimation has gained attention in recent years due to its potential applications in ergonomics, human-robot collaboration systems, and wearable devices. The primary objective of this dissertation was to develop a machine learning based method that can recognize different levels of lifting weight without prior knowledge or measurement of the weight. This research investigated the effects of lifting weights on the upper-body joint kinematics and shoulder-elbow coordination. The results showed that lifting weight level had statistically significant effects on upper extremity kinematics such as velocities and shoulder-elbow coordination. Building on these results and findings from previous studies, a BiLSTM-Transformer Encoder model was developed using lifting kinematics as input features for lifting weight recognition. This model demonstrated improvements over existing multi-level lifting weight recognition models. Furthermore, with feature importance analysis, we found that the model was able to focus on different features during different phases of the lifting process, highlighting its nuanced understanding of the lifting biomechanics involved. Further, the dissertation explored the application of the developed model to video data for lifting weight recognition. This approach demonstrated reasonable results in multi-level lifting weight recognition using video footage. This research demonstrated the potential for improved accuracy in lifting weight recognition tasks. The methods and models developed in this research have promising applications in various areas. The proposed method and findings in this research pave the way for future innovations and developments in lifting weight recognition tasks.