Implementations and evaluation of machine learning algorithms on a microcontroller unit for myoelectric prosthesis control

Abstract: Using a microcontroller unit to implement different machine learning algorithms for myoelectric prosthesis control is currently feasible. Still there are hardware and timing constraints that need to be accounted for. This master thesis presents results from automatically generated Arduino code for some Neural Networks, Convolution Neural Networks and linear machine learning models that was implemented on a Teensy 4.1 board to see where these constraints were on this specific board. The results show promise that simpler algorithms can be generated for 50 classes with an accuracy of around 40-50\%, but more complex algorithms usually run into memory constraints or timing constraints. The results also show that different algorithms are more accurate for different subjects in the used NinaPro database \cite{Atzori2014}. This suggests that configuring the prosthesis on a patient basis, like the one implemented, is useful.