Multi-feature fusion using neural networks for underwater acoustic signal processing

Two neural net (NN) architectures (single NN and multilayer NN), each performing feature fusion for detection and classification of underwater transient signals, are compared. The impact of the different architectures on the training policies is considered; the results of various training schemes are presented, and an attempt is made to obtain an optimal training policy for each of the competing architectures. It is seen that in some cases a neural net using a single input feature can perform well. However, the results obtained indicate that the fused network is the most robust with respect to different levels of signal with additive noise and across the classes considered.<
>