Synergistic Methodology Integrating Flow Noise Numerical Simulation and VRNN Adaptive Denoising for Low-Frequency Hydroacoustic Detection in Underwater Gliders

Maofa Wang Baochun Qiu ^* Zhenjing ZHU Guangtao Shang Zefei Zhu

• Hangzhou Dianzi University, Hangzhou, China

The impact of underwater glider flow noise on low-frequency hydroacoustic target detection constitutes a significant challenge in marine acoustics. To tackle this issue, we employ a hybrid numerical simulation approach combining Large Eddy Simulation (LES) with Lighthill acoustic analogy theory to characterize flow noise profiles around a glider platform. Our analysis identifies two key characteristics: (1) broadband spectral dispersion of flow noise with prominent low-frequency components, and (2) irregular temporal fluctuations in these low-frequency bands. Building on these findings, we propose a novel adaptive noise reduction framework integrating Variational inference with Recurrent Neural Network (VRNN) architecture. This method establishes probabilistic relationships between noise and target signals through latent variable modeling, enabling effective signal-noise separation without prior assumptions about noise characteristics. Experimental validation using four distinct hydroacoustic target classes from the ShipEar dataset demonstrates the framework's superior performance compared to conventional denoising techniques. The VRNN method achieves an average signal-to-noise ratio (SNR) improvement of 11.9 dB, which significantly enhances detection in the critical 0-100 Hz frequency range, and achieves a 98.2% success rate in detecting hydroacoustic targets after noise reduction. This methodology bridges numerical simulation hydroacoustics with data-driven signal processing, offering a deployable solution for next-generation underwater gliders requiring high-fidelity low-frequency surveillance.