Application of attention-based transformer for ship-of-opportunity spectrogram prediction

Passive sonar is a useful tool for underwater acoustics that can be used to detect ships or other sound sources in the ocean. The input signals into the sonar system in turn can also be used to make inferences about the ocean environment. In recent work, ship noise has been used to infer seabed information from 15–20 minute spectrograms, with CPA in the center of the time window. Waiting for 15–20 minutes to record the full time window slows down real-time inference efforts. In this work, our goal is to complete ship-of-opportunity (SOO) spectrograms given the first few minutes. Our approach is to use a self-supervised attention-based transformer, which has been found to be effective for machine learning problems, particularly for natural language processing. The transformer is trained on simulated SOO spectrograms and tested on spectrograms from the Seabed Characterization Experiment in 2017 in the New England Mud Patch. The resulting predicted spectrograms contain the key features of the measured spectrograms over the full time window. If successful, this work may allow for real-time applications of ship detections and seabed inferencing methods. [Work supported by the Office of Naval Research, Grant N00014-22-12402.]