Associative feedforward neural networks with no hidden layers were applied to the problem of localizing a source in range and depth using the acoustic signal arriving at a vertical array of sensors. A highly processed form of the signal (excitations of an orthogonal basis) was used as input in order to increase the robustness of the trained network. The output layer consisted of one unit for each possible range and one unit for each possible depth of the source. The networks were trained with a signal-to-noise ratio (S/N) at the hydrophone of 50 dB, and then their performance was evaluated with S/Ns of 50 and 0 dB. Network weights were found for narrow and broad target shapes that correspond to narrow and broad beamshapes. The narrow target produced the beamformer with the lowest sidelobes and highest gains with acceptable but somewhat higher sensitivity to noise. Performance in the region for which the network was trained compared favorably with minimum variance beamforming.
No takes yet. Share an insight, caveat, or question.
Ozard et al. (1991) studied this question.