Ice anthropogenic classification with acoustic vector sensors using transformer neural networks

Abstract

Acoustic classifiers are a necessary component in understanding the source. When a foreign object has been classified, physics models can be associated with the foreign object for better localization and tracking. In highly non-linear environments, like shallow ice environments, traditional classifiers cannot properly consider its compounded non-linearities: multi-path, reflective surfaces, scattering fields, and the dynamic acoustic properties of first-year ice. With such significantly distorted signals, we deploy deep neural networks to better classify different acoustic sources. We collected data from 8 different acoustic sources on the Keweenaw Waterway in Houghton, Michigan: a narrow and shallow channel covered with first-year ice. Two sources were moving and the other five were stationary; the sources did not emit simultaneously. Data were recorded using two spatially separated underwater acoustic vector sensors; their time-series data were post-processed into mel-frequency cepstral coefficients (MFCC) and analyzed with different deep neural network architectures. A deep Transformer neural network and a deep residual neural network were then compared in their ability to predict which source was emitting. Preliminary results show success with the deep Transformer neural networks.