Graph Attention Network Enhanced Reinforcement Learning for Collaborative USV Patrolling

In addressing the challenge of collaborative patrolling among multiple Unmanned Surface Vehicles (USVs), this study proposes a reinforcement learning-based collaborative patrolling algorithm utilizing information maps. First, the patrolling environment model and unmanned vehicle motion model are established for scenarios involving collaborative patrolling in USV swarm. Second, considering the patrolling history of all vehicles in the swarm, a global information map of the sea area is constructed using radar detection data and the motion trajectories of all patrolling vehicles. Concurrently, to address the obstacle avoidance requirements of the USV swarm, high-precision local environmental information and safety data are provided through the local information map to ensure safe navigation. Finally, a training network based on the Graph Attention Network (GAT) is designed to make the network fusion results independent of changes in the number and arrangement of vehicles, thus enhancing scalability. Simulation experiments demonstrate the effectiveness of the proposed patrolling algorithm, as evidenced by data analysis of several performance indicators. Notably, the algorithm achieves favorable results in the core metric of patrolling interval time across different numbers of vehicles.