Identification of Underwater Propeller Noise by Low-rank Approximation of Cyclic Spectrum
Lei He, Haiyan Wang, Muhang Zhang · 2018
Identification of ships based on the characteristics of propeller radiated noise is a valuable and challenging task. In this paper, we use cyclic spectrum as the propeller fingerprint. The cyclic spectrum can take advantage of various detailed information of the propeller noise, especially the second-order statistics. We improve the identification accuracy by the low-rank approximation of tensors constructed from the cyclic spectrum. Compared with the traditional acoustic signal processing, our method is free of feature extraction and feature optimation. The useful information for identification is extracted from the cyclic spectrum tensor constructed from the training samples. In the meanwhile, the high-dimensional redundancy in the cyclic spectrum and irrelevant information between samples are reduced. We tested the identification method on the propeller noises of two vessels in a well-controlled situation. An accuracy above 95% is obtained with only 1/20 of the samples used as training data. This result indicates that the proposed method has a good application prospect.