Unveiling persistent structures in long-range acoustic m-sequence data with regularized machine learning

Monitoring ocean dynamics via long-range acoustic transmissions is a topic of much interest. Characterizing complex physical systems, such as the ocean, is challenging because their dynamics often cannot be fully described by simple analytical models. Recently, machine learning methods have gained traction in the sonar signal processing community, offering alternatives to classical statistical and signal processing methods. These traditional methods often struggle in real ocean environments, where nonlinear effects, such as internal waves, distort transmitted acoustic pulses. This work investigates using modern machine learning techniques to unravel these complexities by identifying persistent, slowly evolving features. We assess the capabilities of the proposed machine learning methods using transmitted maximum length sequence signal pulses from the Kauai Beacon source collected at the International Monitoring Station near Wake Island. Additionally, we demonstrate that incorporating regularization constraints enhances interpretability, providing deeper insight into the learned models. [Work funded by DoD Navy (NEEC) Grant No. N00174201001, and the ONR grant numbers N000142112420 and N000142312503.]