Minimum entropy filtering for improving nonstationary sonar signal classification

The passive sonar classification problem can be decomposed into two stages: (l) recovering the source time signature of a transient event from a set of received signals by accounting for environmental distortion effects, and (2) applying a pattern recognition algorithm to the estimated source signature for final classification. The minimum entropy method is studied with regard to its performance in removing multipath distortion from passive transients, to improve the performance of classifiers. It was found that the method often works well if the kurtosis of the associated multipath Green's function is high enough, and that signal stationarity is not required. We also found that, while there are usually a few filter lengths at which the best solutions are obtained with conventional convergence criteria, good solutions exist across a much broader range of filter lengths if the iterations are not allowed to proceed to convergence. That is, kurtosis needs to be increased, but not maximized. In many cases, two or three iterations is sufficient.