Automated acoustic arrival matching using a convolutional neural network approach informed by statistics of acoustic scattering from internal waves

Abstract

A machine learning model was developed to automatically align underwater acoustic measurements taken at various depths and ranges from a transmitting source in the Philippine Sea to a reference model of long range acoustic arrival structure, simultaneously determining source-receiver range and travel-time offsets associated with multipath arrivals. Ocean sound-speed variability complicates the task as the measured arrivals may exhibit scattering not present in range-independent predictions. Monte Carlo style broadband parabolic equation simulations through random internal wave fields consistent with the Garrett-Munk internal wave energy spectrum were used to generate a large data set of simulated acoustic receptions including scattered multipath arrivals with known source-receiver ranges and imposed travel time offsets. These simulated receptions were used to train and evaluate the machine learning model for arrival pattern matching to the reference model. The inclusion of various data dimensions, such as peak amplitude and width, and contextual information, such as range and depth, were also explored as input to the model. Ranging results for the machine learning model were compared to a programmatic solution engineered for the same task.