Developpement des criteres d'apprentissage pour le controle d'un bras robot manipulateur a 7 DDL par le traitement des signaux EMG chez les blesses medullaires

The control of a robotic aid poses a great challenge to disabled people who must cope with limited physical ability. A myoelectric control presents an interesting and intuitive solution because of its possible adaptation to most physical limitations. This paper evaluates the use of linear discriminant analysis (LDA) as a tool for multiple limb motion classification using continuous myoelectric signals on high level spinal cord injury (SCI) subjects. The purpose of this work is to create a 16-commands myoelectric controller using generic learning and analysis criteria. To this end, an experimental evaluation of the system was conducted on 23 subjects (12 healthy and 11 SCI). This study evaluates the influence of multiple learning and analysis parameters on the definition of movements which achieve higher classification rate. This work also proposes movements and new communication strategies, likely to be used for control of robotic manipulator.;It is shown that the subject's condition did not influence the results of the classifier. In this study, it was possible to achieve accuracy of 95% on five classes (5) classification using a generic parameterization for learning and analysis of signals. An accuracy of 90% was also achieved for nine class (9) classification problem whereas 82% accuracy for thirteen classes classification problem. The preferred movements for an EMG controller are unilateral elevation of the shoulders and flexion of the elbow. The preferred communication strategies are movements of 2 and 4 seconds, maximum amplitude movements and series of 2 "clicks". Our study also concludes this kind of interface is undesired by SCI population and that their electrical wheelchair joystick is preferred.;Keywords: Electromyography, feature extraction, linear discriminant analysis, spinal-cord injury.