Dependent-Gaussian-Process-Based Learning of Joint Torques Using Wearable Smart Shoes for Exoskeleton

Abstract

Estimating the joint torques of lower limbs in human gait is a highly challenging task and of great significance in developing high-level controllers for lower-limb exoskeletons. This paper presents a dependent Gaussian process (DGP)-based learning algorithm for joint-torque estimations with measurements from wearable smart shoes. The DGP was established to perform data fusion, and serves as the mathematical foundation to explore the correlations between joint kinematics and joint torques that are embedded deeply in the data. As joint kinematics are used in the training phase rather than the prediction process, the DGP model can realize accurate predictions in outdoor activities by using only the smart shoe, which is low-cost, nonintrusive for human gait, and comfortable to wearers. The design methodology of dynamic specific kernel functions is presented in accordance to prior knowledge of the measured signals. The designed composite kernel functions can be used to model multiple features at different scales, and cope with the temporal evolution of human gait. The statistical nature of the proposed DGP model and the composite kernel functions offer superior flexibility for time-varying gait-pattern learning, and enable accurate joint-torque estimations. Experiments were conducted with five subjects, whose results showed that it is possible to estimate joint torques under different trained and untrained speed levels. Comparisons were made between the proposed DGP and Gaussian process (GP) models. Obvious improvements were achieved when all DGP r2 values were higher than those of GP.