PF-LSTM Reinforcement Learning Enhanced Hybrid Acoustic-Optical Adaptive Collaborative AUV Localization Algorithm

Abstract

Cooperative localization of autonomous underwater vehicles (AUVs) is widely used in fields such as ocean exploration and environmental monitoring. However, its effectiveness highly depends on precise position estimation and clock synchronization mechanisms. Clock offset and drift, signal multipath attenuation, and dynamic interference in the underwater medium significantly constrain localization accuracy in anchor-free environments. Although existing cooperative algorithms have proposed solutions like TDOA/TDOC to address asynchrony, they still face challenges such as error accumulation, slow convergence, and energy consumption imbalance. To this end, this paper proposes a Potential Field-LSTM reinforced hybrid acoustic-optical adaptive AUV cooperative localization algorithm (PF-LSTM-QHAACL). The algorithm introduces a reinforcement learning decision framework incorporating Potential Field-based dense reward and LSTM temporal memory modules, thereby accelerating the learning process and improving localization accuracy. Simultaneously, to tackle clock asynchrony dynamics and acoustic-optical channel fluctuations, PF-LSTM-QHAACL employs a DQN-like mode switching mechanism for real-time channel assessment and adaptive training, further optimizing system stability and energy utilization. Furthermore, the algorithm integrates a hybrid ranging strategy combining Time Difference of Arrival (TDOA) and Received Signal Strength (RSS), effectively suppressing the impact of asynchronous bias on position estimation. Simulation results demonstrate that the PF-LSTM-QHAACL algorithm significantly enhances underwater localization accuracy and success rate in highly asynchronous scenarios.