A new strategy optimisation method for underwater flapping foil propulsion based on Twin-Delayed Deep Deterministic and Gaussian process regression

Abstract

In complex sea environments, underwater bionic flapping foils exhibit significant nonlinear characteristics across a high-dimensional parameter space. This study introduces a novel optimisation paradigm, G-TD3, integrating Twin-Delayed Deep Deterministic (TD3) reinforcement learning with Gaussian process regression (GPR). An automatic experimental device was developed, using which a series of flapping foil propulsion experiments were conducted, exploring both symmetric and asymmetric flutter modes. Utilizing GPR, we characterized the flapping foil's operational environment from experimental data. In addition, the TD3 algorithm was deployed to exploit this environment, enabling interactive learning in high-dimensional, continuous parameter spaces. Our integrated approach facilitates comprehensive exploration-exploitation dynamics. The optimized propulsion strategy, fine-tuned through maximizing acquisition rewards, demonstrates G-TD3's superiority over traditional brute-force methods, requiring less data for optimising both symmetric and asymmetric flutter strategies. This research provides a new optimisation method for nonlinear systems and offers valuable insights into the practical applications of bionic flapping foils.