Examinator: Max Ortiz Catalán
Handledare: Jan Zbinden
Opponent: Julia Molin
The most common technique to control active prostheses is via electromyography (EMG), where myoelectric signals from voluntary muscle contractions are recorded and interpreted as different movements. One longstanding challenge in EMG prosthetic control is crosstalk. Where crosstalk in EMG can be explained as the part of an EMG signal recorded over a muscle but produced by another muscle. The appearance of crosstalk makes decoding of the intended prosthetic movement more difficult and the resulting control less reliable which in turn affects the performance of the prosthesis.
This project investigates how PCA and ICA affect the EMG signals acquired from implanted electrodes of two participants, one with a trans-humeral and one with a trans-radial amputation. Classification with an MLP showed that the application of PCA over offline featured EMG signals could increase the accuracy for pattern recognition from 87.7% to 99.9%. Meanwhile, the application of ICA increased the accuracy from 87.7% to 93.6%. One modified offline method with PCA which also could be used online was tested and resulted in an increased accuracy from 87.7% to 93.9%. If this improved performance could be achieved online these findings could in the long term improve prosthetics control and yield better life quality for people suffering from limb loss.