Advanced Myocontrol for Hand and Wrist Prostheses

Abstract

Myocontrol is the use of a human machine interface based on muscle signals in order to control a robotic or prosthetic device. A challenging problem in research is the simultaneous and proportional (s/p) control of multiple degrees of freedom (DOF). Besides the common sensing technique of surface electromyography (sEMG), force myography (FMG) is additionally enforced to improve the control experience. Machine learning approaches are employed to map the input of different sensor modalities onto continuous control signals of a prostheses. Throughout this work, four experiments have been conducted with the goal to reduce the training time and to improve the control experience of such devices. In the first experiment we showed that for a fusion of both signal modalities the online performance is invariant for different sensor placements on the forearm. The second experiment evaluated the online performance using different machine learning approaches where either one or both signal modalities were employed. The best results were achieved with a combination of both signal types and with FMG only. As part of this work, an existing method called linearly enhanced training (LET) is adapted to the multi-modal sensory input. This method creates artificial training data for combinations of defined hand and wrist actions and dismisses their explicit recording as training data, which usually cannot be achieved by amputees. It follows that the training time is significantly reduced. In the related experiment, data has been gathered from 10 healthy subjects in order to find a generalized set of parameters for LET. Once determined, these parameters can be the basis for LET for new users. In the last experiment, the set of generalized parameters has been used for nine healthy subjects to evaluate the performance of the approach involving LET data. We showed that with LET the subjects performed equally well compared to the approach which required the execution of all combined activations during training time. This qualities LET as a valid extension to existing control methods as the training time is drastically reduced and no combined activations need to be executed. The goal is to use the same set of parameters and algorithm for amputees, which may not be able to produce combined activations during training time.