Model Predictive Control Using Linearized Neural Network-Based NARX Modeling for Antagonistic Pneumatic Artificial Muscle Systems

This paper presents a data-driven modeling and control framework for antagonistic pneumatic artificial muscle (PAM) systems, aiming to achieve accurate joint trajectory tracking in rehabilitation applications. A Multi-Layer Feedforward Neural Network (MFNN) is embedded in an NARX structure to model the nonlinear system dynamics and is locally linearized to obtain a computationally efficient prediction model. Based on this model, a constrained model predictive control (MPC) scheme is developed, explicitly considering physical limits on the control increment, control input, and system output. An online disturbance estimation is incorporated into the prediction model to enhance robustness against external loads and unmodeled dynamics. The resulting constrained quadratic program is efficiently solved using the Alternating Direction Method of Multipliers (ADMM), enabling real-time implementation. Experimental results on a PAM-driven knee joint, under both sinusoidal and human gait trajectories, demonstrate improved tracking accuracy, smoother control actions, and guaranteed constraint satisfaction compared to unconstrained MPC, confirming the effectiveness of the proposed approach for rehabilitation robotics.