Towards precise and rapid gearbox noise prediction: a physics-informed region-aware acoustic predictor

Abstract Noise is a crucial dynamic performance indicator for gearboxes. However, direct and precise measurement of actual vibration and noise during gearbox operation under complex working conditions is challenging and obtaining noise performance by directly measuring prototype samples during the design phase is cost-prohibitive. Traditional numerical analysis methods struggle to meet real-time measurement requirements. While emerging purely data-driven methods, due to their black-box nature, suffer from a lack of reliability and generalization. This paper introduces a revolutionary computational paradigm to overcome these limitations: the physics-informed region-aware acoustic predictor (PI-RAAP). The key innovation of PI-RAAP is a physically-motivated architecture that explicitly models the underlying physics of acoustic radiation. Within a dual-encoder design, a novel region-aware mechanism mathematically encodes the acoustic superposition principle. This structure enables the model to quantitatively attribute the contribution of each surface region and thus provides much improved interpretability. Furthermore, the governing Helmholtz equation is embedded in the loss function as a soft constraint, which helps to maintain physical consistency and improves generalization beyond the training data. A comprehensive numerical simulation pipeline is established and experimentally validated for generating the dataset. Results demonstrate that PI-RAAP achieves exceptional predictive accuracy, evidenced by an average SPL error of just 0.702 dB, while accelerating prediction speeds by at least 3 orders of magnitude compared to conventional BEM. The proposed framework strongly promotes the in-depth application and widespread adoption of digital twin technology in the analysis and control of gearbox vibration and noise by accurately substituting actual measurement with real-time prediction.