Information decoupling via comparative filtering: A robust fault diagnosis framework for rotating machinery in high-noise environments

Abstract

Fault diagnosis of rotating machinery is severely hampered by heavy background noise, which obscures fault signatures and degrades model performance. Conventional deep learning methods attempt to address this by enhancing noise resistance, but they often fail to distinguish between noise and fault characteristics, leading to suboptimal results. To overcome this limitation, this paper introduces a novel fault diagnosis framework based on information decoupling. Our approach fundamentally separates the noise reduction task from the diagnosis task. It first employs a systematic search algorithm to identify and construct a representative noise profile from the raw signal itself. Subsequently, a comparative filtering module, featuring a reverse attention mechanism, utilizes this profile to suppress noise components in the frequency domain. Finally, a dedicated fault diagnosis network analyzes the denoised signal to achieve accurate classification. Extensive experiments on public and proprietary datasets validate the framework’s superiority. Notably, under extremely challenging conditions (e.g., −9 dB SNR on a helicopter transmission shaft dataset), the proposed method achieves a diagnostic accuracy of 89.16%, surpassing the performance of recent advanced models by a significant margin of over 10 percentage points, demonstrating its superior robustness for real-world applications.