Learning-Enhanced Online Fault-Tolerant Replanning Method for Multi-AUV Systems in Dynamic Flow Fields

:Objectives This study investigates online fault-tolerant replanning methods for multi-autonomous underwater vehicle (AUV) formations operating in dynamic and uncertain flow fields. These formations are prone to task chain interruptions and system utility degradation caused by sudden faults such as power warnings, hardware damage, and communication failures. Methods A long short-term memory (LSTM) network is used to perceive the temporal uncertainty of the flow field and to construct a risk assessment mechanism. A proximal policy optimization (PPO) algorithm dynamically adjusts the computational budget and task retention strategy according to the system state. A graph neural network (GNN) explores the structural correlations between tasks and AUVs to guide evolutionary search. These modules operate in coordination with a warm-started anytime evolutionary framework to achieve online recovery and continuous optimization under limited time budgets. Results Simulation results demonstrate that under the default scenario, the proposed method achieves an AUV survival rate of 89% and a critical task completion rate of 80%. Compared with the ablation models without PPO decision-making and without GNN guidance, the critical task completion rate is improved by 36 percentage points and 35 percentage points, respectively. The minimum average system recovery time reaches 11 seconds, enabling rapid fault recovery and continuous task execution under limited time budget constraints.Conclusions By integrating three learning mechanisms—risk perception, adaptive resource scheduling, and structure-guided search—the proposed method effectively improves the online fault-tolerant replanning performance of multi-AUV systems under dynamic ocean currents and unexpected failures, providing a useful reference for autonomous fault tolerance and collaborative decision-making in complex marine environments.