Robust Anomaly Detection of Rotating Machinery with Contaminated Data

Abstract

Rotating machinery is critical to industrial systems, necessitating robust anomaly detection (AD) to ensure operational safety and prevent failures. However, in real-world scenarios, monitoring data is typically unlabeled and often consists of normal samples contaminated with a small proportion of unknown anomalies. To address this, this paper proposes a diffusion-based AD method, Anomaly Detection Denoising Diffusion Probabilistic Model(AD-DDPM) for robust anomaly detection. The method employs a U-attention-net to capture local and global features and introduces a filtered contrastive mechanism to mitigate the impact of contaminated training data. By leveraging the probabilistic nature of diffusion models, AD-DDPM effectively models normal data distributions, achieving superior anomaly detection even with polluted samples. Experimental validation on fault simulation datasets demonstrates the method’s exceptional performance, outperforming traditional machine learning and deep learning baselines. The proposed approach offers a promising solution for reliable health monitoring in industrial settings.