Combined separation and classification of two types of coexistent ship radiated noise based on trained ideal ratio mask and cepstral features

It is common that the radiated noise samples recorded by underwater hydrophone are mixture of radiated noises from different ships. When dealing with multichannel data, beamforming technique can be used to separate different sources from different directions. However, the resulting signal may still contain other sources because of the resolution limitation and sidelobe leakage. So methods on separating the radiated noise from different ships on time-frequency domain are expected, which will have significant impact on the following classification task. In this work, on an actual measurement database, a multi-layer perceptron network was trained for estimating ideal ratio masks (IRM) for both ships on Mel spectrogram and then Mel cepstral features extracted from the separated Mel spectrogram were used for classification. On an actual measurement database of two ships in which most samples are mixed samples, instead of discarding the mixed samples, the proposed system can make use of more samples to build a more powerful classifier with improved generalization performance.