Deep Learning Based Upper-limb Motion Estimation Using Surface Electromyography

To advance human-machine interfaces (HMI) that can help disabled people reconstruct lost functions of upper-limbs, machine learning (ML) techniques, particularly classification-based pattern recognition (PR), have been extensively implemented to decode human movement intentions from surface electromyography (sEMG) signals. However, performances of ML can be substantially affected, or even limited, by feature engineering that requires expertise in both domain knowledge and experimental experience. To overcome this limitation, researchers are now focusing on deep learning (DL) techniques to derive informative, representative, and transferable features from raw data automatically. Despite some progress reported in recent literature, it is still very challenging to achieve reliable and robust interpretation of user intentions in practical scenarios. This is mainly because of the high complexity of upper-limb motions and the non-stable characteristics of sEMG signals. Besides, the PR scheme only identifies discrete states of motion. To complete coordinated tasks such as grasping, users have to rely on a sequential on/off control of each individual function, which is inherently different from the simultaneous and proportional control (SPC) strategy adopted by the natural motions of upper-limbs. The aim of this thesis is to develop and advance several DL techniques for the estimation of upper-limb motions from sEMG, and the work is centred on three themes: 1) to improve the reliability of gesture recognition by rejecting uncertain classification outcomes; 2) to build regression frameworks for joint kinematics estimation that enables SPC; and 3) to reduce the degradation of estimation performances when DL model is applied to a new individual. In order to achieve these objectives, the following efforts were made: 1) a confidence model was designed to predict the possibility of correctness with regard to each classification of convolutional neural networks (CNN), such that the uncertain recognition can be identified and rejected; 2) a hybrid framework using CNN for deep feature extraction and long short-term memory neural network (LSTM) was constructed to conduct sequence regression, which could simultaneously exploit the temporal and spatial information in sEMG data; 3) the hybrid framework was further extended by integrating Kalman filter with LSTM units in the recursive learning process, obtaining a deep Kalman filter network (DKFN) to perform kinematics estimation more effectively; and 4) a novel regression scheme was proposed for supervised domain adaptation (SDA), based on which the model generalisation among subjects can be substantially enhanced.