Energy Reduction for Wearable Pneumatic Valve System With SINDy and Time-Variant Model Predictive Control

Abstract

Pneumatic actuators are a popular choice for wearable robotics due to their high force-to-weight ratio and natural compliance, which allows them to absorb and reuse wasted energy during movement. However, traditional pneumatic control is energy inefficient and difficult to precisely control due to nonlinear dynamics, latency, and the challenge of quantifying mechanical properties. To address these issues, we developed a wearable pneumatic valve system with energy recycling capabilities and applied the sparse identification of nonlinear dynamics (SINDy) algorithm to generate a nonlinear delayed differential model from simple pressure measurements. Using first principles of thermal dynamics, SINDy was able to train time-variant delayed differential models of a solenoid valve-based pneumatic system and achieve good testing accuracy for two cases—increasing pressure and decreasing pressure, with training accuracies at 85.23% and 76.34% and testing accuracies at 87.66% and 77.66%, respectively. The generated model, when integrated with model predictive control (MPC), resulted in less than 5% error in pressure control. By using MPC for human assistive impedance control, the pneumatic actuator was able to output the desired force profile and recycle 85% of the energy used in negative work. These results demonstrate an energy-efficient and easily calibrated actuation scheme for designing assistive devices such as exoskeletons and orthoses.