Abstract
 

Accurate reconstruction of vibration signals from rotating machinery is essential for reliable fault diagnosis and condition monitoring. In practical industrial environments, however, strong and unpredictable background noise often obscures critical fault-related features, substantially degrading the accuracy and robustness of diagnostic systems. Despite recent advances, conventional deep learning-based denoising methods frequently fail to sufficiently suppress complex noise or tend to distort essential signal components, particularly under highnoise conditions. These limitations are especially problematic in scenarios where early fault detection is crucial for preventing equipment failures and minimizing downtime. To address these challenges, we propose a novel signal reconstruction framework based on a denoising diffusion process. The proposed method introduces two key innovations. First, a frequency-focused attention mechanism selectively emphasizes meaningful spectral components, ensuring that fault-relevant frequency characteristics are preserved. Second, a reverse step prediction module adaptively estimates the optimal reverse diffusion step for each input, preventing over-processing and maintaining signal fidelity. This dual approach enhances both frequencydomain consistency and denoising performance throughout the reconstruction process. Extensive experiments conducted on open-source rotating machinery datasets demonstrate that the proposed framework effectively restores essential signal components and preserves fault-relevant information, even under severe noise conditions. These results confirm the superior performance and robustness of our method compared to conventional deep learning-based denoising techniques, highlighting its practical applicability to real-world industrial scenarios.