Detection and characterization of ship underwater radiated narrowband noise

Abstract

Underwater radiated noise (URN) emitted by ships is a high-intensity noise that interferes with acoustic transmissions. It is composed of broadband cavitation noise, tonal noise and its harmonics that can be observed up to several tens of kHz. The in-situ detection of ship noise and the identification of its tonal components are therefore useful for adaptive underwater communication techniques and for channel sensing. To this end, we present a detection scheme to identify the presence of a vessel and identifying its narrowband components. For presence detection, we develop a Convolution Neural Network (CNN) model whose input is a Detection Envelope Modulation On Noise (DEMON) analysis for a given observation window. The model is trained with our collected recordings of ship data and ambient noise. After identifying the ship’s URN, our tonal detector relies on the stability and stationarity of the ship’s tonal lines, as opposed to the randomness of the ambient noise. Cross-correlation and spectral entropy are used as detection metrics. To reduce the sensitivity to the tested environment, the detection thresholds are set adaptively. The results show a favorable trade-off between precision and recall compared to benchmark methods. We share our database of URN of labeled ships.