Underwater Source Localization Based on Dynamic Weight Meta-Learning

Underwater source localization by deep neural networks is challenging because training neural networks requires a large amount of experimental data and high computational costs. In this paper, a dynamic weight meta-learning method (DWML) for underwater source localization is proposed, where the meta-training stage is trained using tasks constructed from simulation data, and the meta-testing stage is for the localization task of experimental data. Specifically, a dynamic weight adjustment function is designed to allocate adaptive weights for different tasks and combined with the label distribution vector to supervise the training of the localization model. Experimental results show that the proposed DWML significantly improves the localization performance with very limited experimental data, and the localization performance is further improved after training with the label distribution vector.