Robust Electromyography Based Control of Multifunctional Prostheses of The Upper Extremity
von Sebastian Amsüss
Datum der mündl. Prüfung:2014-09-26
Betreuer:Prof. Dr. Florentin Wörgötter
Gutachter:Prof. Dr. Florentin Wörgötter
Gutachter:Prof. Dr. Dr. Dario Farina
EnglischMultifunctional, highly dexterous and complex mechanic hand prostheses are emerging and currently entering the market. However, the bottleneck to fully exploiting all capabilities of these mechatronic devices, and to making all available functions controllable reliably and intuitively by the users, remains a considerable challenge. The robustness of scientific methods proposed to overcome this barrier is a crucial factor for their future commercial success. Therefore, in this thesis the matter of robust, multifunctional and dexterous control of prostheses of the upper limb was addressed and some significant advancements in the scientific field were aspired. To this end, several investigations grouped in four studies were conducted, all with the same focus on understanding mechanisms that influence the robustness of myoelectric control and resolving their deteriorating effects. For the first study, a thorough literature review of the field was conducted and it was revealed that many non-stationarities, which could be expected to affect the reliability of surface EMG pattern recognition myoprosthesis control, had been identified and studied previously. However, one significant factor had not been addressed to a sufficient extent: the effect of long-term usage and day-to-day testing. Therefore, a dedicated study was designed and carried out, in order to address the previously unanswered question of how reliable surface electromyography pattern recognition was across days. Eleven subjects, involving both able-bodied and amputees, participated in this study over the course of 5 days, and a pattern recognition system was tested without daily retraining. As the main result of this study, it was revealed that the time between training and testing a classifier was indeed a very relevant factor influencing the classification accuracy. More estimation errors were observed as more time lay between the classifier training and testing. With the insights obtained from the first study, the need for compensating signal non-stationarities was identified. Hence, in a second study, building up on the data obtained from the first investigation, a self-correction mechanism was elaborated. The goal of this approach was to increase the systems robustness towards non-stationarities such as those identified in the first study. The system was capable of detecting and correcting its own mistakes, yielding a better estimation of movements than the uncorrected classification or other, previously proposed strategies for error removal. In the third part of this thesis, the previously investigated ideas for error suppression for increased robustness of a classification based system were extended to regression based movement estimation. While the same method as tested in the second study was not directly applicable to regression, the same underlying idea was used for developing a novel proportional estimator. It was validated in online tests, with the control of physical prostheses by able-bodied and transradial amputee subjects. The proposed method, based on common spatial patterns, outperformed two state-of-the art control methods, demonstrating the benefit of increased robustness in movement estimation during applied tasks. The results showed the superior performance of robust movement estimation in real life investigations, which would have hardly been observable in offline or abstract cursor control tests, underlining the importance of tests with physical prostheses. In the last part of this work, the limitation of sequential movements of the previously explored system was addressed and a methodology for enhancing the system with simultaneous and proportional control was developed. As a result of these efforts, a system robust, natural and fluent in its movements was conceived. Again, online control tests of physical prostheses were performed by able-bodied and amputee subjects, and the novel system proved to outperform the sequential controller of the third study of this thesis, yielding the best control technique tested. An extensive set of tests was conducted with both able-bodied and amputee subjects, in scenarios close to clinical routine. Custom prosthetic sockets were manufactured for all subjects, allowing for experimental control of multifunction prostheses with advanced machine learning based algorithms in real-life scenarios. The tests involved grasping and manipulating objects, in ways as they are often encountered in everyday living. Similar investigations had not been conducted before. One of the main conclusions of this thesis was that the suppression of wrong prosthetic motions was a key factor for robust prosthesis control and that simultaneous wrist control was a beneficial asset especially for experienced users. As a result of all investigations performed, clinically relevant conclusions were drawn from these tests, maximizing the impact of the developed systems on potential future commercialization of the newly conceived control methods. This was emphasized by the close collaboration with Otto Bock as an industrial partner of the AMYO project and hence this work.
Keywords: EMG; Pattern Recognition; Prosthesis; Patient Evalauations; Machine learning; Robustness