Artificial neural networks for simultaneous and independent range and depth discrimination in passive acoustic localization

Two feedforward neural networks with one hidden layer each were trained using a modified backpropagation algorithm to determine the position of an acoustic source in a waveguide. One network was trained to localize the source in depth while the other was trained independently to localize in range. The signal was preprocessed by decomposition along an orthogonal basis vector set in order to increase the robustness of the resulting trained network to uncertainties in the signal and environmental parameters. The output layer consisted of one unit for each possible range or depth of the source. The networks were trained with a signal-to-noise ratio (S/N) of 50 dB and tested with patterns generated with S/Ns of 50 dB and 0 dB. Unambiguous localization was achieved with the trained network at 50 dB S/N, but the localization was more sensitive to the added noise at 0 dB S/N than a perceptron trained with one output cell for each combination of range and depth