Improved signal detection in non-Gaussian deep ocean noise.

Abstract

DeepSound is an autonomous, high-bandwidth acoustic recording system designed to profile ambient noise to depths of 9 km. Recent ambient noise measurements recorded by the DeepSound probe well below the conjugate depth of the channel exhibit significant non-Gaussianity (the hypothesis of Gaussianity is rejected for 97% of 5 s samples using the large sample size corrected Anderson–Darling test at a 95% confidence level). A common sonar signal processing task is the automatic detection of a signal corrupted by ambient ocean noise. The standard solution is matched filtering, which is the optimal detector in the maximum likelihood sense assuming additive white Gaussian noise. Because matched filter is derived under the assumption of Gaussian noise, its performance suffers when implemented in real world non-Gaussian noise. This paper describes a signal detection algorithm for non-Gaussian noise based on the generalized Gaussian probability density function. The generalized Gaussian detector is locally optimal for small SNR, and the processing chain resembles a matched filter with a non-linear preprocessing stage. The detection performance of the generalized Gaussian detector is shown to exceed that of matched filtering for synthetic signals injected into the noise measurements recorded by the DeepSea probe. [Work supported by ONR.]