Model-guided Bayesian U-net for vessel trajectory estimation using cepstrum

In the near-field, the cepstrum of broadband ship radiation noise can exhibit the time difference of arrival (TDOA) between two dominant propagating paths in an oceanic waveguide, which contains the information about ship’s kinematics. This study proposes a model-guided deep learning framework that employs a Bayesian U-Net–based architecture to automatically infer shipping information such as ship trajectory, ship speed, closest approach point, and more. To overcome the lack of reliable and high-quality training data, the network is trained solely with a large variety of synthetic training datasets obtained from the passive sonar simulator, while it is evaluated using real measurements collected from coastal waters around Korea, as well as publicly available datasets. The results show that the learned representation effectively reconstructs the shipping information from real measurements. Finally, model uncertainty and data uncertainty are quantitatively analyzed using the proposed Bayesian framework. [Work supported by the Korea Institute of Marine Science & Technology Promotion (KIMST) funded by the Ministry of Oceans and Fisheries (RS-2023-00256122).]