Towards Robust Regression-based Machine Learning for Myoelectric Control

This thesis examines robustness of regression algorithms for myoelectric control against undesired disturbances and proposes new approaches to enhance the robustness. The analyzed disturbances include force variation in electromyography signals, subjects' arm position change, and subjects' learning ability. Both offline and real-time experimental results validate that the proposed methods could be possible solutions to narrow the gap between a controlled laboratory environment where myoelectric control has been often tested and the real-world environment where non-stationary disturbances are present. Overall, this work contributes to the development of myoelectrical prostheses with regression algorithms more usable to amputees.