A Feature Self-Supervised Learning Method Driven by DEMON Spectral Model for Underwater Noise of Ships

The existing intelligent underwater acoustic target recognition mainly relies on a large amount of data to train deep neural network models. However, due to the difficulty of obtaining data in the underwater acoustic field and the scarcity of labeled measured data, the training data conditions are not sufficient, making it difficult to obtain recognition models with generalizability. This article constructs a physical model driven underwater acoustic DEMON(demodulation on noise)spectral feature self-supervised learning model and provides corresponding training methods. By using physical models and network self-supervised processes to constrain the direction of network learning, the problem of lots of training data demand in deep learning is alleviated. Finally, the performance of the model was verified through simulation data experiments. The accuracy of the DEMON spectrum blade number and shaft frequency extracted by the method in this paper is as high as 100%, which is higher than or equal to the accuracy of 97.82% and 100% extracted using only neural networks.

I. Introduction

Underwater noise of ships target recognition is a crucial step in sonar signal processing. Underwater acoustic target recognition mainly relies on the analysis and processing of ship radiated noise. The radiation noise of ships is mainly composed of mechanical noise, propeller noise, and hydrodynamic noise[1]. Due to the complex and ever-changing underwater environment, various interfering sound sources, as well as the attenuation and frequency offset generated during the propagation process, can cause distortion of ship radiated noise signals, which poses difficulties for the classification and recognition of underwater acoustic targets. The identification of radiated noise from underwater target ships is an urgent problem to be solved.