Ankle Torque Estimation Using HDEMG Driven CNN-LSTM Model and Data Augmentation

Sammanfattning: Robotic-powered exoskeletons are increasingly used to assist patients with movement disorders in daily life and rehabilitation. Accurately estimating joint torque, especially for dynamic movement conditions using EMG, is crucial for effective assistance. Machine learning and deep learning have been employed for EMG-based force/torque estimation, but their precision and robustness have been limited, particularly for dynamic movements. This thesis aims at comparing and analyze the results using MLP, CNN, and CNN-LSTM methods to estimate ankle joint torque in dynamic movements based on HD-EMG. Meanwhile, this thesis designed and tested different data augmentation to enhance the performance using HD-EMG data augmentation techniques. The CNN-LSTM model demonstrated superior performance among the machine learning models. Additionally, the combination of spatial and signal augmentation methods showed notable improvements in the inter-subject case performance of the prediction. No augmentation methods have shown notable improvements in the intra-subject case or inter-session case.