The fundamental frequency, or F0, is directly determined by the vibration frequency of sound source. Therefore, by detecting and tracking F0, it is possible to estimate the attributes and motion state of an underwater maneuvering object, including the rotation speed of the propeller, the shaft number and so on. Unfortunately, coupled with mechanical noise and continuous noise, and further polluted by the marine ambient noise, some spectral lines that correspond to the the fundamental frequency and its harmonics would be distorted, like amplitude attenuation or position shift, which severely discounts the detection precision of the state-of-art methods. To improve the precision and stability of F0 detection, especially in low SNR situation, we propose to estimate the state of sound source by learning the hydrophone array signal with a deep neural network. Specifically, in the preprocessing stage, the DEMON spectrum is extracted from noise signal for each hydrophone channel, and it is subsequently cleaned by the wavelet denoising. Then, we use a comb filter to compensate the spectral line distortion effects. Finally, the purified frequency spectrum features are fed into a CNN network to determine F0 by classification. Simulation experiments show that the data-driven deep learning method outperforms traditional model-based algorithms in the case of low-SNR noise signals.

This study is published in 2020 3rd International Conference on Unmanned Systems (ICUS).