Integrated models of signals and background for an HMM/neural net ocean acoustic event classifier

The authors investigate the use of hidden Markov models (HMMs) for the classification and detection of ocean acoustic events in a nonstationary ocean background. A statistical formalism is described for integrating models for dynamic acoustic events and ocean background into a unified statistical framework. In this framework, both signal processes and background processes are modeled as HMMs, and signal classification is performed by obtaining the likelihood of a corrupted observation sequence through a combined state space of signal and background. Techniques are presented for estimating the acoustic event model parameters from training exemplars that are observed in these difficult background conditions. A novel neural network technique is proposed for the automatic learning of the nonlinear mechanism through which signal and background observations interact. Experimental results are presented