A Study of Submersible Positioning Based on Deep Reinforcement Learning and Automatic Control Algorithms

In this paper, a position prediction model and a search model are developed to explore the application of Monte Carlo modeling and deep reinforcement learning and control algorithms to the localization of marine submersibles. A set of six-degree-of-freedom spatial equations of motion are developed using Newton's second law for analyzing the motion of a submersible in still water. Simulations confirm a high degree of similarity between the model and the actual motion. The Monte Carlo model is used to simulate the probability distribution, optimize the initial deployment position, and improve the efficiency of repositioning the faulty submersible. Among the compared search methods, reinforcement learning shows higher efficiency and resource savings with a minimum relative search time of 0.57 and a search completion rate of 0.95. An automatic control system combining deep reinforcement learning and control algorithms provides a hierarchical framework for real-time search path optimisation with improved adaptability and efficiency. For the multi-submarine task, an improved k-means clustering algorithm is used to construct a dynamic partitioning model to optimize the configuration of finite search devices. Simulation results show that the probability-based deployment scheme achieves a higher search success rate compared to the traditional approach.