A Composite Observer-Based Optimal Attitude Tracking Control for FWEPAUV via Reinforcement Learning

The foldable wave-energy powered autonomous underwater vehicle (FWEPAUV) is capable of directly generating sufficient electrical energy from seawater when its body aligns perpendicular to the wave flow direction. Its precise attitude control is of vital importance in long-term navigation missions. This article presents a neural network-based adaptive optimized attitude control approach for an FWEPAUV system, considering unmeasurable yaw angular velocity, system uncertainties, and external disturbances. The composite observer and radial basis function neural networks (RBF NNs) are established to estimate the yaw angular velocity, system uncertainties, and disturbances. The estimated signals are then jointly utilized to compensate for the redundant signals in the control channel, thereby enhancing robustness. A composite-observer-actor-critic reinforcement learning architecture is proposed to learn the optimal value function, generate the control torque, and achieve the balance between the control accuracy and cost. The introduced prescribed performance mechanism ensures the smoothness of the FWEPAUV's transient response and the accuracy of the steady-state response while guaranteeing all error signals semi-globally uniformly ultimately bounded (SGUUB). The effectiveness of the developed approach is exemplified via a case study.