Integrated front reconstruction and AUV tracking control with Bayesian optimization and NMPC

Abstract

Ocean fronts are the boundaries where distinct water masses meet. This process involves energy exchange and nutrient transport that both are of significant interest for the marine research. This paper considers using an Autonomous Underwater Vehicle (AUV) to sample, plan, control, and reconstruct the front autonomously. To achieve the task, an integrated reconstruction, path planning, and control framework is proposed with Bayesian optimization and nonlinear model predictive control (NMPC). The integration lies in that the performances of front reconstruction and AUV path tracking influence each other. Particularly, differing from the traditional predetermined or rule-based sampling pattern, a Bayesian optimization scheme is proposed without defining a temperature detection threshold for ocean fronts. During each iteration of the Bayesian optimization, the Gaussian process regression is performed first to reconstruct the temperature field of the front region in a data-driven way. Next, an acquisition function characterizing the possible location of the front is maximized to generate a series of waypoints. Those waypoints are further converted into Zig-zag-like smooth reference paths of the AUV. This ensures that the vehicle moves in the region of interest. Then, with the line-of-sight guidance law, we derive the AUV error dynamics model based on which an NMPC controller is proposed unifying the online planning and control considering system constraints explicitly. The goal is to achieve high path-tracking accuracy which also means high-value front regions can be visited by the AUV. Monto Carlo simulations are carried out with high-fidelity data from regional ocean models. It is illustrated that the AUV deployed in the front region can accomplish the autonomous sampling, reconstruction, and tracking tasks with satisfactory efficiency and precision. The proposed method has superior reconstruction and tracking accuracy performance than several classic methods.