Influence of acoustic field interference structure on underwater acoustic target recognition based on a convolutional neural network

Abstract

The acoustic field in a shallow sea waveguide has a complex spatial interference structure caused by the sea surface and seabed [1]. This spatial interference will distort the spectrum of the target signal in the propagation process, increasing the difficulty of recognizing underwater acoustic targets. Of course, sufficient target data at different receiving positions can improve the generalization ability of a classifier model. This solution can partly arrest the degradation of the recognition rate, but the high cost and difficulty of acquiring experimental data offset this advantage. Instead, researchers have begun expanding the simulation target data using sound propagation models [2-4]. The target data obtained by this method are reliable. Sufficient simulation data can also compensate the insufficient experimental data. In shallow ocean with low frequency, the normal mode (NM) model can accurately and quickly calculate the acoustic field. Therefore, the NM model is often chosen as the acoustic field calculation model. The NM model represents acoustic field by the superposition of various normal modes. Although the data expansion method based on ocean acoustic fields can improve the recognition rate of targets, which normal modes dominate the recognition rate improvement remains unclear. And classifier based on the deep learning model has high generalization ability, partial acoustic field interference may not affect the recognition rate of targets. When identifying the normal modes that mainly influence the recognition rate, target recognition can be simplified, which has considerable significance.