Identification of Vessels from Engine Sounds by Spectral Comparison and Verification

We consider the problem of identifying a vessel from its engine sound. A database of known vessels is used for comparison of acoustic characteristics in the frequency domain. Only one training sample is required from each class for classification. With more training samples, a vessel is not only classified but also verified whether it belongs to the same class as its closest match in the database. The acoustic signals are regarded as stationary time series with mixed spectra. We boost their power by adding a small amount of noise in order to assist classification. Three nonparametric estimators of spectra are compared in classification and verification experiments. The case of a vessel not being represented in the database is also considered with the intention to have it rejected at the verification stage.