A deep neural network approach to acoustic source localization in a shallow water tank experiment

In this paper, an acoustic source localization method using the emerging technology of the deep neural network (DNN) is proposed. After the construction and training of the DNN, the capability of the DNN for source localization through a set of numerical simulations is verified. Next, experimental studies and demonstrations in a very shallow water tank with acoustic reflective walls are prepared, which enable the quick acquisition of a huge amount of experimental data for the training of a one-dimensional DNN-based source localization model. The development of the DNN-based source localization method and the corresponding numerical and experimental demonstration constitute the main contribution of this work. The associated performance is then evaluated at various frequencies. In particular, the localization results of the DNN are compared with readily available model-based localization methods, such as the conventional matched field processing method and the normal-mode based multiple signal classification method. The comparison shows that the proposed DNN approach is able to produce satisfactory accuracy in this reflective shallow water tank environment, for which a forward acoustic propagating model is not required. Last but not least, the generality of the proposed DNN approach from one-dimensional localization to progressively more complicated two-dimensional tasks is also considered.