A Long Short-Term Memory-Based Predictive Control Approach for Pneumatic Muscle Control Systems

Abstract

Pneumatic artificial muscles (PAMs) are essential in robotics systems designed to mimic the human musculoskeletal structure, particularly in applications for rehabilitation, due to their flexibility and muscle-like force output. However, their nonlinear and hysteresis characteristics challenge accurate modeling and limit traditional model-based control effectiveness. This study presents a hybrid framework combining long short-term memory (LSTM) networks with model predictive control (MPC), significantly improving PAM control precision. The approach ensures reliable trajectory tracking without model switching, supported by a novel stability theorem. Experimental results reveal that the LSTM-MPC method outperforms conventional controllers, achieving improvements in root-mean-square error of 0.0051, 0.0312, and 0.0125 compared to proportional–integral–derivative, adaptive fuzzy control, and Learning MPC, respectively, while reaching the reference trajectory within 3 s. Load-bearing tests on the pneumatic knee joint demonstrated effective adaptation of system inputs for trajectory tracking under step and sinusoidal signals, and achieve commendable performance in knee joint motion tracking. This research offers a new perspective on PAM control, emphasizing the integration of deep learning with traditional control methods.