Improved time-frequency de-noising of acoustic signals for underwater detection system

Abstract

The capability to communicate and perform target localization efficiently in underwater environment is important in many applications. Sound waves are more suitable for underwater communication and target localization because attenuation in water is high for electromagnetic waves. Sound waves are subjected to underwater acoustic noise (UWAN), which is either man-made or natural. Optimum signal detection in UWAN can be achieved with the knowledge of noise statistics. The assumption of Additive White Gaussian noise (AWGN) allows the use of linear correlation (LC) detector. However, the non-Gaussian nature of UWAN results in the poor performance of such detector. This research presents an empirical model of the characteristics of UWAN in shallow waters. Data was measured in Tanjung Balau, Johor, Malaysia on 5 November 2013 and the analysis results showed that the UWAN has a non-Gaussian distribution with characteristics similar to 1/f noise. A complete detection system based on the noise models consisting of a broadband hydrophone, time-frequency distribution, de-noising method, and detection is proposed. In this research, S-transform and wavelet transform were used to generate the time-frequency representation before soft thresholding with modified universal threshold estimation was applied. A Gaussian noise injection detector (GNID) was used to overcome the problem of non-Gaussianity of the UWAN, and its performance was compared with other nonlinear detectors, such as locally optimal (LO) detector, sign correlation (SC) detector, and more conventionally matched filter (MF) detector. This system was evaluated on two types of signals, namely fixed-frequency and linear frequency modulated signals. For de-noising purposes, the S-transform outperformed the wavelet transform in terms of signal-to-noise ratio and root-mean-square error at 4 dB and 3 dB, respectively. The performance of the detectors was evaluated based on the energy-to-noise ratio (ENR) to achieve detection probability of 90% and a false alarm probability of 0.01. Thus, the ENR of the GNID using S-transform denoising, LO detector, SC detector, and MF detector were 8.89 dB, 10.66 dB, 12.7dB, and 12.5 dB, respectively, for the time-varying signal. Among the four detectors, the proposed GNID achieved the best performance, whereas the LC detector showed the weakest performance in the presence of UWAN.