posted on 2015-11-19, 08:59authored byGareth. Loudon
This thesis relates to the use of knowledge based signal processing techniques in the decomposition of EMG signals. The aim of the research is to fully decompose EMG signals recorded at fairly high force levels (up to twenty percent maximum voluntary contraction) automatically into their constituent motor unit potentials to provide a fast and accurate analysis routine for the clinician. This requires the classification of non-overlapping motor unit action potentials (MUAPs) and superimposed waveforms formed from overlapping MUAPs in the signal. Firstly, digital filtering algorithms are used to reduce noise in the signal. A normalisation and compression of the filtered signal is then performed to reduce the time of the analysis. Non-overlapping MUAPs are classified using a statistical pattern recognition method. The method first describes the MUAPs by a set of features and then uses diagonal factor analysis to form uncorrelated factors from these features. An adaptive clustering technique groups together MUAPs from the same MU using the uncorrelated factors. The decomposition of superimposed waveforms is divided into two sections. The first section is a procedural method that finds a reduced set of all possible combinations of MUAPs which are capable of forming each superimposed waveform. The second section is a knowledge based analysis of the selected MUAP combinations forming each superimposed waveform. An expert system has been designed to decide which combination is the most probable by studying the motor unit firing statistics and performs uncertainty reasoning based on fuzzy set theory. The decomposition method was tested on real and simulated EMG data recorded at different levels of maximum voluntary contraction. The different EMG signals contained up to six motor units (MUs). The new decomposition program decomposed all MUAPs in the EMG signals tested into their constituent MUs with an accuracy always greater than ninety five percent. The decomposition program takes about fifteen seconds to classify all non-overlapping MUAPs in an EMG signal of length one second and on average, an extra nine seconds to classify every superimposed waveform. Hardware limitations did not enable the testing of EMG signals containing more than six MUs. The results also show that the computer analysis can simulate the reasoning of a human expert when studying a complex EMG signal.