Machine learning for direction of arrival estimation using an underwater glider mounted acoustic compact volumetric array with experimental results

The paper deals with the estimation of the direction of arrival (DOA) of an acoustic signal measured using a compact volumetric array using a machine learning approach. The approach consists on using Gaussian Process Regression (GPR) for training and subsequently using the obtained model to estimate the DOA. The input of the GPR algorithm are the phase differences of the signal measured by the different hydrophones and optionally the sign of the pitch of the vehicle. Using the sign of the pitch improves the result. The approach is validated using an experimental dataset consisting of a source sending wide band chirps and a glider equipped with the 6 hydrophone compact volumetric array system (cVAS). The glider was performing helicoidal trajectories which provides enough angular diversity for the estimation process to work. The machine learning approach is compared to the classic geometric approach for acoustic arrays. A significantly higher Median Absolute Deviation (MAD) is observed on the elevation estimates when using the geometric approach. The machine learning algorithm is capable of filtering some acoustic effects which cannot be easily modeled using a geometric approach. The MAD of the azimuth are comparable for all three approaches.