Multitask convolutional neural network for acoustic localization of a transiting broadband source using a hydrophone array

A multitask convolutional neural network (CNN) is trained to localize the instantaneous position of a motorboat throughout its transit past wide aperture linear array hydrophones located 1 m above sea floor in water 20 deep. cepstrogram database for each hydrophone and cross-correlogram pair adjacent are compiled multiple transits. Cepstrum-based correlation-based feature vectors (along with ground-truth source bearing range data) form inputs train three CNNs so that they can predict other “unseen” It shown operating on multi-sensor cepstrum-based maps able transiting motorboat, even when near an endfire direction. Also, generalised cross presence interfering multipath arrivals. When compared cepstrum-only CNN, correlation-only conventional model-based method passive ranging by wavefront curvature, combined cepstrum-cross correlation CNN provide superior localization performance underwater acoustic environment.