Title: Neural Radiated-Noise Fields for Unmanned Underwater Vehicle Noise Spectrum Prediction in Three-Dimensional Scenes

Abstract:

Radiated noise serves as a critical metric for defining the acoustic signature and assessing the performance capabilities of unmanned underwater vehicles (UUVs). Traditional approaches, which rely heavily on physics-based modeling and numerical simulations, often struggle with their intense requirement for detailed target structural data and specific environmental boundary conditions. Furthermore, these conventional methods fail to provide continuous spatial spectrum-response modeling within three-dimensional environments. To overcome these limitations, this study introduces the Neural Radiated-Noise Field (NRNF).

The NRNF approach models the radiated-noise spectrum of a UUV as a continuous function that accounts for three-dimensional UUV positioning, three-dimensional hydrophone placement, the vehicle’s yaw angle, and signal frequency. This framework allows for prediction queries at any arbitrary spatial location. The methodology utilizes sinusoidal encoding for positional and frequency inputs and incorporates a learnable three-dimensional scene feature grid. This grid is designed to explicitly capture environmental structures and acoustic propagation effects.

The model was trained and evaluated using a dataset derived from lake trials. Performance was assessed across three distinct scenarios: horizontal extrapolation, depth extrapolation, and cross-run generalization. The findings indicate that the NRNF maintains an average prediction error of 3.5 dB across the 50 to 5000 Hz frequency band. In terms of difficulty, horizontal extrapolation proved to be the most straightforward task, while depth extrapolation presented the greatest challenge, with cross-run generalization falling in between. Additional ablation studies confirm that the inclusion of the scene feature grid substantially enhances both the spatial generalization capabilities and the overall prediction stability of the model.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC