Classification of EMG signals to control a prosthetic hand using time-frequesncy representations and Support Vector Machines.

摘要

Myoelectric signals (MES) are viable control signals for externally-powered prosthetic devices. They may improve both the functionality and the cosmetic appearance of these devices. Conventional controllers, based on the signal's amplitude features in the control strategy, lack a large number of controllable states because signals from independent muscles are required for each degree of freedom (DoF) of the device. Myoelectric pattern recognition systems can overcome this problem by discriminating different residual muscle movements instead of contraction levels of individual muscles. However, the lack of long-term robustness in these systems and the design of counter-intuitive control/command interfaces have resulted in low clinical acceptance levels. As a result, the development of robust, easy to use myoelectric pattern recognition-based control systems is the main challenge in the field of prosthetic control.;This dissertation addresses the need to improve the controller's robustness by designing a pattern recognition-based control system that classifies the user's intention to actuate the prosthesis. This system is part of a cost-effective prosthetic hand prototype developed to achieve an acceptable level of functional dexterity using a simple to use interface. A Support Vector Machine (SVM) classifier implemented as a directed acyclic graph (DAG) was created. It used wavelet features from multiple surface EMG channels strategically placed over five forearm muscles. The classifiers were evaluated across seven subjects. They were able to discriminate five wrist motions with an accuracy of 91.5%. Variations of electrode locations were artificially introduced at each recording session as part of the procedure, to obtain data that accounted for the changes in the user's muscle patterns over time. The generalization ability of the SVM was able to capture most of the variability in the data and to maintain an average classification accuracy of 90%.;Two principal component analysis (PCA) frameworks were also evaluated to study the relationship between EMG recording sites and the need for feature space reduction. The dimension of the new feature set was reduced with the goal of improving the classification accuracy and reducing the computation time. The analysis indicated that the projection of the wavelet features into a reduced feature space did not significantly improve the accuracy and the computation time. However, decreasing the number of wavelet decomposition levels did lower the computational load without compromising the average signal classification accuracy.;Based on the results of this work, a myoelectric pattern recognition-based control system that uses an SVM classifier applied to time-frequency features may be used to discriminate muscle contraction patterns for prosthetic applications.