Abstract
In the direct arrival zone of the deep ocean, the multi-path time delays have been used for acoustic source localization. One of the challenges in conventional localization methods is to artificially determine which paths the extracted delays belong to. A convolutional neural network, taking the autocorrelation functions as the input feature directly, is proposed for source localization to avoid the path determination procedure. Since some multi-path arrivals may not be visible due to absorption in the bottom of the ocean, a data augmentation method based on a ray propagation model is proposed. Tests on simulated and real data validate the method.
Highlights
Machine learning (ML) has been used in underwater acoustic source localization,1–14 direction-of-arrival estimation,15 and seabed classification.16,17 Passive source localization is posed as a machine learning problem using input features derived from conventional signal processing methods
The training errors of convolutional neural networks (CNNs)-3, autocorrelation function (ACF)-NET, and ACF-NETþ are shown in Figs. 3(a) and 3(b)
The mean absolute error (MAE) versus signal-to-noise ratio (SNR) of (c) depth and (d) range estimates of ACF-NETþ with legends indicating the deleted multi-path arrivals
Summary
Machine learning (ML) has been used in underwater acoustic source localization, direction-of-arrival estimation, and seabed classification. Passive source localization is posed as a machine learning problem using input features derived from conventional signal processing methods. Single-hydrophone source localization can be achieved by exploiting the ocean acoustic propagation attributes such as the interference patterns and multi-path arrival structures.. The capability of the conventional model-based methods, matching the multi-path arrival structures (e.g., time delays) from a single hydrophone, have been demonstrated for range and depth estimation. The time delays obtained from the ACF have been used for moving source localization or source ranging.21 One challenge of these model-based methods is to manually determine which paths the time delays belong to. Inspired by the conventional model-based methods, the ACFs are used as input features to train the neural network This method does not need the path determination procedure (i.e., the multi-path time delays are not extracted artificially).
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
