Novel Machine Learning Methods for Robust Myoelectric Control

点击翻译按钮获取中文摘要

My research group has conducted seminal work in developing pattern recognition-based control of powered artificial limbs using the electrical signals produced by contracting muscles (the myoelectric signal). We remain world leaders in algorithm development and assessment of performance. The primary contributions have come through novel developments and machine learning including feature and classifier design, and the characterization of the impact of training, motor learning and feedback on real-time control. This research has formed the basis of successful commercialization of embedded myoelectric control systems, with the formation of a company by two of my former graduate students. While work remains to improve the reliability and dexterity of control for artificial limbs, there is tremendous potential for innovative use of the myoelectric signal as a human-machine interface (HMI) for existing and emerging consumer applications. An appealing aspect of the myoelectric signal is that it is generated as a consequence of natural muscle contraction and, if movement intent is reliably extracted, it may be used to autonomously control devices. With sensors placed on the arms, its use as a generalized HMI allows use heads-up control scenarios and, with able-bodied users, leaves the hands free for other tasks. This has many potential applications in virtual and augmented reality systems, and auxiliary control during navigation (cycling or driving a vehicle). Although considerable knowledge may be translated from the great success in prosthetics control, optimal design for consumer applications will require novel development of machine learning algorithms, sensor design, and training methodologies. This research program will focus on optimizing the dexterity and robustness of myoelectric control as a generalized human computer interface. These goals will be accomplished by developing novel methods of information extraction and leveraging motor control and learning. My long-term objective envisions developing signal processing and training strategies that will be sufficiently flexible and scalable to be used in future sensing technologies, including peripheral nerve and cortical implants. My short-term objectives are:  1.Develop analytic methods that explicitly incorporate and leverage temporal information to enable new control modalities and improve reliability.  2.Use deep learning methods to improve resilience to distortion.  3.Develop optimized feature and classifier design for electrodes that reside inside the muscles.  4.Develop powerful visual feedback methods to induce motor learning to induce motor learning during training that will result in better real-world performance. This work will provide engaging multidisciplinary training opportunities for students and develop highly sought-after skillsets in problem solving, critical thinking, signal processing, machine learning, and human machine interaction.