System-informed neural network for solving linear inverse problem in underwater acoustics

In underwater acoustics, various applications, such as frequency detection, direction of arrival (DOA) estimation, and localization, can be formulated as linear inverse problems. High-performance solvers for linear inverse problems are essential for the detection and tracking of marine objects such as surface ships and submarines. In this study, we propose a two-step training framework, comprising pre-training and fine-tuning, to implement a system-informed neural network (SINN) solver based on the system model of linear inverse problems, ensuring adaptability to various environments. The SINN architecture adopts an adaptive learned iterative shrinkage thresholding algorithm (Ada-LISTA) of linear inverse problem solver. To overcome the challenge of collecting a large volume of high-quality underwater acoustic data, we generate simulation data that reflects the physical characteristics of the system model and use it to train the model during the pre-training step. In the subsequent fine-tuning step, the model is retrained using a limited amount of underwater acoustic data, with the pre-trained weights serving as initial weight values. This two-step approach facilitates rapid adaptation of the model to the target environment under analysis. The SINN solver not only demonstrates superior restoration and denoising performance but also offers the advantage of reduced computational processing time.