Evolving Gaussian Process Autoregression Based Learning of Human Motion Intent Using Improved Energy Kernel Method of EMG

Abstract

Continuous human motion intent learning may be modeled using a Gaussian process (GP) autoregression based evolving system to cope with the unspecified and time-varying motion patterns. Electromyography (EMG) signals are the primary input. GP is used as a mathematical foundation to model human kinematics by adopting the nonlinear autoregressive with exogenous inputs (NARX) framework, and an evolving system is applied to learn the irregular and unspecified dynamic features. The statistical nature of the GP offers superior flexibility for learning human kinematics and is capable of giving credibility to motion intent prediction, which also enables risk-based control. As an important neuromuscular signal, EMG is processed with a novel method, the energy kernel method, to extract the activation level of muscle and feature out muscular force and motion intent. Without losing robustness, the high signal-to-noise ratio or the linearity level with muscular force, huge improvement has been made in computational efficiency to meet the requirements of real-time applications. Experimental works concerning the validity and application of this method are also presented.