Master thesis presentation – Effect of bioelectrical signal acquisition on classification performance
Presented by Pontus Lövinger
Date: Thursday, May 22nd
Time: 14.00 pm
Location: Landahlsrummet (room 7430), Hörsalsvägen 9, 7th floor, Chalmers
Examiner: Bo Håkansson
Myoelectric robotic prosthetic devices are assistive devices for people with amputations. These devices can be controlled by training a myoelectric pattern recognition (MPR) algorithm with myoelectric signals (MES) sampled from electrodes placed over muscles. Conventionally the MPR algorithms have been train with MES data produced when performing a movement with sustained contraction force. In this thesis, the effect of training the algorithms with the contraction force increasing as a ramp over time has been investigated. For this, the offline accuracies as well as real-time accuracies have been analyzed in order to evaluate the performance of classifiers trained with ramp or sustained contraction data. Two pattern recognition algorithms (LDA and MLP) have been compared for real-time classification. From the offline accuracy it was found that the average accuracy is higher when training classifiers with sustained contraction data compared to ramp contraction data. The average accuracy for real-time classification was found to be similar (no statistically significant difference) when comparing classifiers trained with sustained and ramp contraction data. Of the two algorithms compared for real-time classification using ramp data, linear discriminant analysis (LDA) was found to be better than multilayer perceptron (MLP). From the data it was not possible to find any connection between the offline and real-time accuracies except that the offline accuracies are high than the real-time accuracies.