Abstract:
Exciting new approaches have led to a great increase in the ability of prosthetic devices to accurately replicate the functionality of the human hand. Advanced methods of neural interfacing exist, however they are expensive and not commonly available. Electromyographic (EMG) pattern recognition techniques exist to allow determination of user intention, however they rely on prior knowledge of muscle groups, use a symmetrical array to place electrodes or a simple reduction algorithm, and often require large amounts of data to accurately classify user intention. The aim of this project was to provide improved accuracy of myoelectric prosthetic classification system, without inhibiting its ability to operate in real-time. In order to be able to investigate the affect of various parameters, rich data sets were acquired from the forearm of 10 subjects, through creation of a silicone armband with embedded electrodes. Current methods have had success using a statistical cluster analysis measure on a non-amputee subject, with an eye on future applications to real world prosthetic wearers. This thesis attempts to improve prosthetic control accuracy through the attainment of three goals. First, a real-time capable preprocessing step was performed. A method has been presented to extract action potentials from forearm surface EMG signals, and to have this extracted action potential used in a matched filtering approach. The algorithm was able to identify action potential signals, with average positive predictive value (number of true positives as a percent of total positive results) of 97%. Total processing time involved for this method was less than 10% the length of a signal segment, proving that it is able to be used in real-time applications. This increased the signal-to-noise ratio by, on average, 2.2 dB for data from 10 subjects. Further analysis however showed that this technique was not helpful in improving accuracy in prosthetic control, and due to success with other methods it was not pursued further. The second goal was to provide a means to choose optimal data sets from a larger, irregularly spaced grid of data. This was realized through an interpolation scheme followed by a heuristic data set replacement algorithm. When tested on a reduced number of data sets, this algorithm performed to within 99.9% of an exhaustive algorithm, while taking 1 X 10-5 of the time. Finally, these methods were applied to forearm surface EMG signals in order to compare the performance of the optimized sites to symmetrically placed sites, and two previously used location selection algorithms. For the data processed off-line, this showed a great increase in classification accuracy from 85% to approximately 95%.