Model Compensation with RBF Network based Nonlinear MPC and its Application on an Antagonistic Pneumatic Artificial Muscle System

Tackling the prevalent challenge of unknown model elements and perturbations in practical systems poses a significant barrier to enhancing control precision. This paper proposes a novel RBF-based Nonlinear MPC for mode compensation. Initially, the conventional approach of dynamic modeling is utilized to identify and isolate unmodeled characteristics. Subsequently, Radial Basis Function (RBF) neural networks are employed to predict and compensate for these unmodeled parts. Driven by the sampled data, this method efficiently explores the control action space to improve control performance. Our three-layer neural network architecture significantly reduces computational overhead, and online network updates effectively mitigate neural network generalization issues. We apply the proposed approach to force tracking control of Antagonistic Pneumatic Artificial Muscles (APAM) in flexible structures. Case studies demonstrate a significant improvement in control accuracy compared to the feedforward PID control method.