GDSAN: A physics-informed generative network for zero-shot labeled fault data generation
Cheng Hu , Tielin Shi , Jiantao Lu , Jianqiang Liang , Bin Jia , Jian Duan
Aero-engines are critical industrial assets whose failures can lead to severe consequences, highlighting the necessity of effective Prognostics and Health Management (PHM). However, existing approaches suffer from limitations in data availability and model accuracy, particularly when real fault samples are scarce or absent, hindering reliable diagnostics. This study develops a novel physics-informed network, named Generative Data-simulation Adversarial Network (GDSAN), to generate labeled fault vibration signals for reliable aero-engine rotor systems health monitoring. This model introduces a learnable modifying matrix to systematically reconcile discrepancies between simulated and measured data across four error dimensions. After that, physics-informed spectral and energy constraints are embedded into the loss function to enhance both model training stability and the physical plausibility of generated signals. Furthermore, a hybrid-driven PHM framework is constructed, leverages former generated labeled fault data to realize zero-shot fault diagnosis, thereby reducing reliance on high-fidelity simulation models or extensive measured fault samples. The following experimental validation on an aero-engine test bench demonstrates that the proposed framework successfully generates labeled fault signals closely aligned with experimental measurements in both the feature space and frequency spectrum, and eliminates the desperate need for enormous but expensive measured fault samples in model training process. Moreover, the proposed physics-informed terms in the loss function significantly improve the physical plausibility of generated signals.
aero-engine / rotor system / data generation / physics-informed / hybrid-driven / prognostics and health management
Higher Education Press 2026
/
| 〈 |
|
〉 |
PDF
(7820KB)
