Deep learning approach to generate propeller cavitation noise

Abstract

In underwater acoustic simulators, propeller cavitation noise has traditionally been modeled as a modulated broadband signal. This study aims to enhance the realism of these simulators by employing deep learning to generate propeller cavitation noise. The training data were collected from the modeled propeller under various pressure conditions in the Samsung Cavitation Tunnel. We have utilized a variant of Generative Adversarial Networks (GANs), wherein both the generator and discriminator are designed with a recursive structure. To assess the advantages of our data-based approach, we analyze the characteristics of the signals generated and compare them with the real signals or the modeled signal based on physics. The results demonstrate the effectiveness of our proposed model in generating realistic propeller cavitation signals. [Work supported by the Korea Research Institute for Defense Technology Planning and Advancement (20-106-B00-003).]