Continuous Intention Prediction of Lifting Motions Using EMG-Based CNN-LSTM

Abstract

Industrial exoskeletons is a field of ongoing research for improving human safety and conveniences. However, the adoption of industrial exoskeleton robots still remains challenging. One problem that needs to be solved is the control delay that inevitably occurs due to data transmission and processing issues. Recently, there has been active research employing deep learning to address control delay by leveraging diverse information extracted from human motion. One crucial source of information is electromyography (EMG) signals, known for their quicker activation compared to actual motion. This study specifically focused on predicting changing motion intentions within the squat, a representative lifting motion in industrial contexts. In an experimental setup involving 24 participants and utilizing 7 EMG electrodes, we categorized motion intentions during the squat into four types. We developed a CNN-LSTM model capable of predicting motion intentions 300 milliseconds ahead using EMG signals. The model’s prediction performance was assessed by comparing them with existing models. The findings propose a methodology for utilizing EMG signals in predicting changing motion intentions for lower extremity movements, facilitating feedforward control of industrial exoskeleton robots. This research is anticipated to contribute to the advancement of acceptance in the realm of exoskeleton robots.

Keywords

• CNN-LSTM
• exoskeleton
• intention prediction
• deep learning
• EMG
• feedforward control