A Robust Anomaly Detection Model for Pumps Based on the Spectral Residual With Self-Attention Variational Autoencoder

Tianming Xie, Qifa Xu, Cuixia Jiang, Zhiwei Gao, Xiangxiang Wang

Research output: Contribution to journalArticlepeer-review


Multivariate time-series (MTS) collected from multiple sensors on industrial pumps often exhibit concept drift and noise contamination due to variable working conditions and complex environments. To detect anomalies in such MTS, we propose a novel model called spectral residual with self-attention variational autoencoder (SR-SAVAE). Specifically, the spectral residual operation is used to mitigate concept drift, while the variational inference combined with a total variation regularization is used to address the issue of noise contamination. Experimental results on three public datasets indicate that the SR-SAVAE model achieves good anomaly detection results for general MTS. More importantly, compared to other state-of-the-art models on a private dataset about pumps, the results illustrate the superiority of the SR-SAVAE model in anomaly detection for MTS with concept drift and noise contamination. Finally, ablation studies on the SR-SAVAE model detail the efficacy of each component.
Original languageEnglish
Pages (from-to)1-11
Number of pages11
JournalIEEE Transactions on Industrial Informatics
Early online date9 Apr 2024
Publication statusE-pub ahead of print - 9 Apr 2024

Cite this