Passive acoustic localization using dictionary learning

Abstract

Ray-trace simulations are used to demonstrate the use of Dictionary Learning to improve passive localization of a source monitored with a vertical linear acoustic array. The learned dictionary, generated using historical sound velocity profile (SVP) data from a region of interest, is an over complete, sparse representation of the SVP training set. By minimizing an objective function that measures the differences between multipath reception intervals detected at a receiver for propagation through candidate and baseline SVP profiles, we show that an optimal SVP match to the “unknown” baseline can be reconstructed by an efficient dictionary search. Ray traces back-propagated through the reconstructed SVP according to beamformed receive angles computed with the baseline SVP are then found to well estimate the source position as the centroid of a cluster of multipath signal intersections. The accuracy for small unit-sparsity dictionaries of size up to 50 was evaluated on a randomly sampled, 30 SVP testing set obtained from an area close to that of the training set, demonstrating mean location errors less than 5% in both distance and depth. Five representative profiles spanning the range of observed sound speed variations were used to assess localization performance at source depths from 50 to 450 m and at source ranges from 2000 to 4500 m. Compared to using the average sound speed profile to estimate source position, using the learned dictionary produced a general performance increase in position accuracy.