Quantifying bias in underwater radiated noise measurement

Abstract

The measurement of ship underwater radiated noise (URN) is a foundational element of underwater noise management policies. These measurements are frequently made in shallow water (<50 m). Current methods estimate the power spectrum with a single transmission loss applied for all frequencies. A latent assumption in this is signal stationarity, i.e., propagation loss properties do not change with ship position relative to the hydrophone. A breach of the stationarity assumption leads to bias in calculating the average. Using this work's proposed methodology, the stationarity assumption is shown to be false in the along-track dimension for one ship at one site at two different times of the year, from 30 to 300 Hz. This band covers many noise sources of interest on ships. Linear regression methods demonstrate some violations of stationarity but produce smaller estimates of bias. Single layer neural networks are shown to produce higher levels of estimated bias but only marginally better mean absolute error and root mean squared error statistics for an ideal ship model. When neural networks models are applied back to the original run data, it is found the magnitude of real bias lies between that of the regression models and the ideal ship model.