Robust multivariate time series anomaly detection via conditional normalizing flow with patch embedding and cross time attention

As a key technology for the intelligence of cyber-physical systems, multivariate time series anomaly detection has permeated numerous core scenarios in the Industrial Internet of Things (IIoT). However, anomaly detection remains highly challenging due to the relative scarcity of labeled anomalies, class imbalance between normal and anomalous samples, and the diversity of unknown anomalies. Additionally, accurately modeling the complex temporal patterns and multivariate correlations in time series poses another significant difficulty. Furthermore, the widely adopted one-class classification (OCC) methods in this domain are constrained in real-world applications due to their stringent requirement for completely normal training data. To address these challenges, we propose a novel unsupervised multivariate time series anomaly detection method named PANF that integrates P atch embedding and cross time A ttention to enhance temporal patterns extraction within a conditional N ormalizing F low framework. To model critical local semantic information in time series, we utilize patch embedding to map series into a latent state space, capturing rich local temporal dynamics. To address complex temporal dependencies, cross time attention is applied to learn temporal representations and model inter-time correlations. The conditional normalizing flow is employed to estimate time series density for anomaly detection, thereby eliminating OCC's reliance on clean training sets. Experiments on three real-world datasets demonstrate that our method outperforms other mainstream anomaly detection models, proving its superior performance and potential for real-world deployment.