Data-driven approaches to reducing the training burden in pattern recognition based myoelectric control
University of New Brunswick
Advancements in EMG pattern recognition have enabled intuitive gesture recognition of upper limb motions for use in human computer interfaces. Whether it be to control a prosthesis, a drone, or a virtual reality environment, all current implementations rely on the acquisition of representative training data from the end-user before use. This requirement has been reported as frustrating for users and a limitation to the broader adoption of EMG-based controllers. Consequently, this thesis aims to address the burden of the training protocol for EMG pattern recognition using data-driven approaches and cross-user models. Two exploratory studies were conducted to assess the impact and degree of inter-subject variability and difference in EMG information content between user groups. Based on observed subject differences, an adaptive domain adversarial neural network (ADANN) was subsequently developed to adapt a previously trained model to a new user using minimal training data. The proposed ADANN cross-subject model significantly outperformed the current state-of-the-art canonical correlation analysis (CCA) cross-subject model for both intact-limb and amputee populations (with 9.4% and 22.4% absolute improvement, respectively). Finally, a generative adversarial network architecture, SinGAN, was adopted as a novel alternative for reducing the amount of EMG data needed for training. SinGAN was able to generate synthetic EMG signals based on a single subject-supplied motion repetition, significantly improving accuracy compared to training with the single motion alone.