Abstract:
Rapidly increasing developments in the field of artificial intelligence yield implications for motion supportive devices. Powered ankle prostheses (PAP) employ actuation mechanisms that aim to support amputees during locomotion. Surface electromyography (sEMG) allows muscles contribution in developing such algorithms and their usage as the exclusive sensing source is novel. Different techniques of processing raw sEMG data exist, but their optimal usage in a PAP control algorithm has not been elaborated on. Despite their benefits, normalization of the raw sEMG data and usage of sliding window might be problematic for a real time application. The main aim of this thesis was to develop and test control algorithms for a PAP that successfully distinguishes step-up and step-down tasks of healthy population using exclusively sEMG. A specific aim was to assess the effects of normalization and windowing procedures of sEMG data on prediction success of the algorithms developed. An open database was used to acquire sEMG data of 50 participants. Four machine learning techniques namely, (i) ANN, (ii) LSTM, (iii) LightGBM and (iv) RF were used with windowed/non-windowed and normalized-non-normalized datasets. Comparison of their performances showed that LightGBM utilizing non-normalized and non-windowed dataset performed,(Accuracy=92.53%) not significantly different than the best performance obtained utilizing normalized and windowed dataset(Accuracy=94.02%). So, it is concluded that LightGBM can be used in step-up/down classification module of the control algorithm for PAP using non-normalized sEMG data.|Keywords : EMG Signal, Machine Learning, Decision Trees, Deep Neural Networks
