Investigating the efficacy of autoencoders and other machine learning methods for studying dynamic oceanic processes in long-range acoustic propagation environments

Abstract

Long-range acoustic propagation is a topic of great interest in applications like acoustic thermometry and underwater navigation. These applications utilize the measured arrival times and structure obtained from the transmitted probe pulses. However, they are highly dependent on the stability and repeatability of the ocean channel. In real oceans, random medium effects like internal waves [1] can induce considerable fluctuations and distortions to the received probe pulses. Our objective here is to investigate the use of machine learning (ML) methods such as autoencoders and other deep learning architectures to see if they can unravel and give insight into the dynamics of the ocean processes generating the fluctuations. In particular, we will investigate geometric concepts from braid, loops, and knot theory that can capture the changing shapes of smoothly deforming features that represent these processes. For the analysis, we will use transmitted frequency maximum length sequence (MLS) signal probe pulses from the 75 Hz Kauai Beacon source received at the International Monitoring Station near Wake Island at a nominal distance of 3500 km. We show that ML analysis can provide some useful insights. J. Xu, “Effects of internal waves on low frequency, long range, acoustic propagation in the deep ocean,” MIT Ph. D. dissertation (2007).