Feature extraction techniques for grasp classification.

This thesis examines the ability of four signal parameterisation techniques to provide discriminatory information between six different classes of signal. This was done with a view to assessing the suitability of the four techniques for inclusion in the real-time control scheme of a next generation robotic prosthesis. Each class of signal correlates to a particular type of grasp that the robotic prosthesis is able to form. Discrimination between the six classes of signal was done on the basis of parameters extracted from four channels of electromyographie (EMG) data that was recorded from muscles in the forearm. Human skeletal muscle tissue produces EMG signals whenever it contracts. Therefore, providing that the EMG signals of the muscles controlling the movements of the hand vary sufficiently when forming the different grasp types, discrimination between the grasps is possible. While it is envisioned that the chosen command discrimination system will be used by mid-forearm amputees to control a robotic prosthesis, the viability of the different parameterisation techniques was tested on data gathered from able-bodied volunteers in order to establish an upper limit of performance. The muscles from which signals were recorded are: the extensor pollicis brevis and extensor pollicis longus pair (responsible for moving the thumb); the extensor communis digitorum (responsible for moving the middle and index fingers); and the extensor carpi ulnaris (responsible for moving the little finger). The four signal parameterisation techniques that were evaluated are: 1. Envelope Maxima. This method parameterises each EMG signal by the maximum value of a smoothed fitted signal envelope. A tenth order polynomial is fitted to the rectified EMG signal peaks, and the maximum value of the polynomial is used to parameterise the signal. 2. Orthogonal Decomposition. This method uses a set of orthogonal functions to decompose the EMG signal into a finite set of orthogonal components. Each burst is then parameterised by the coefficients of the set of orthogonal functions. Two sets of orthogonal functions were tested: the Legendre polynomials, and the wavelet packets associated with the scaling functions of the Haar wavelet (referred to as the Haar wavelet for brevity). 3. Global Dynamical Model. This method uses a discretised set of nonlinear ordinary differential equations to model the dynamical processes that produced the recorded EMG signals. The coefficients of this model are then used to parameterise the EMG signal 4. EMG Histogram. This method formulates a histogram detailing the frequency with which the EMG signal enters particular voltage bins) and uses these frequency measurements to parameterise the signal. Ten sets of EMG data were gathered and processed to extract the desired parameters. Each data set consisted of 600 grasps- lOO grasp records of four channels of EMG data for each of the six grasp classes. From this data a hit rate statistic was formed for each feature extraction technique. The mean hit rates obtained from the four signal parameterisation techniques that were tested are summarised in Table 1. The EMG histogram provided Parameterisation Technique Hit Rate (%) Envelope Maxima 75 Legendre Polynomials 77 Haar Wavelets 79 Global Dynamical Model 75 EMG Histogram 81 Table 1: Hit Rate Summary. the best mean hit rate of all the signal parameterisation techniques of 81%. However, like all of the signal parameterisations that were tested, there was considerable variance in hit rates between the ten sets of data. This has been attributed to the manner in which the electrodes used to record the EMG signals were positioned. By locating the muscles of interest more accurately, consistent hit rates of 95% are well within reach. The fact that the EMG histogram produces the best mean hit rates is surprising given its relative simplicity. However, this simplicity makes the EMG histogram feature ideal for inclusion in a real-time control scheme