Deep Learning-based Ultra Short Baseline Underwater Positioning

In ultra-short baseline (USBL), the locations of near-field sources are estimated by using the difference between the propagation delays for the received signals of sensors. Since the sensor spacing is very narrow in the USBL, the difference between the propagation delays for the received signals is very small, which induces ambiguities in positioning for the sources. For low sampling rate scenarios with low signal-to-noise power ratios (SNRs), the ambiguities increase significantly because not only the sample delays for the received signals may not be exactly estimated, but also the difference between the sample delays for the received signals decreases. To solve this problem, this paper proposes a deep learning-based USBL positioning network. The inputs of the proposed network are the estimated distances from the source to the sensors, which are measured by cross-correlation, and the outputs are the range and direction-of-arrival (DOA) of the near-field source. The proposed network improves the positioning performances even if outliers, i.e., incorrectly estimated sample delays, are mixed in the input by learning the relationship between the input and output. Computer simulations demonstrate that the proposed network has 50 times better positioning performances than the conventional method in low SNR regions.