Unsupervised Anomaly Detection Based on Diffusion Model‐Enhancement for Bearing Vibration Signals

ABSTRACT

Unsupervised anomaly detection plays a crucial role in prognostics and health management. Industrial equipment predominantly operates in a normal state throughout its service life, rendering faults rare and the collection of sufficient fault samples impractical. This necessitates unsupervised approaches that establish normality boundaries based on normal data. We propose a denoising diffusion probabilistic model (DDPM)‐enhanced anomaly detection framework, which integrates data augmentation techniques based on the DDPM, efficient dimensionality reduction via principal component analysis, and anomaly detection using the Isolation Forest algorithm. The proposed framework is applied to the CWRU and MFPT datasets, and the results demonstrate robust cross‐dataset consistency, with the AUC‐ROC differing by only 0.32 percentage points between the two datasets. Extensive comparative experiments show that DDPM outperforms the variational autoencoder (VAE) in data augmentation, achieving superior interclass separation (8.09σ vs. 5.70σ). Furthermore, early fault detection experiments reveal that under 10% fault signal contamination, DDPM‐based augmentation achieves 82% recall compared to 60% for VAE‐based augmentation. These results indicate that the proposed DDPM framework is effective for industrial condition monitoring, enabling efficient dimensionality reduction without compromising detection accuracy.