Research on Cross-Device Bearing Fault Diagnosis Transfer Method Based on NICE-GAN

In industrial applications, acquiring fault signals under diverse operating conditions remains a significant challenge, particularly for cross-device and cross-condition fault diagnosis. This paper proposes a novel transfer learning framework based on the No-Independent-Component-of-Encoder Generative Adversarial Networks (NICE-GAN) model without independent encoding components to address this challenge. The framework leverages both simulation and experimental data to generate high-fidelity fault signals for target domain diagnosis. Our approach comprises three key components: First, we establish a comprehensive rolling bearing dynamics model based on Hertzian contact theory to generate abundant simulation signals under various working conditions, which are then transformed into time-frequency representations. Second, we employ the NICE-GAN architecture to learn shared features between simulation and experimental data under identical working condition constraints. Third, we utilize the trained generator to synthesize realistic fault data that closely matches experimental characteristics while preserving the underlying fault mechanisms from simulation data. Extensive experimental validation demonstrates that our method achieves diagnostic accuracy approaching that of real experimental data in cross-device fault diagnosis tasks, outperforming existing approaches by a significant margin. This framework provides a promising solution for practical fault diagnosis applications where fault data acquisition is limited.