AD-FGP: Industrial Multivariate Time-Series Anomaly Detection via Fusion of Generative and Predictive Models.

Anomaly detection on industrial multivariate time-series data is an important research topic for industrial control systems. Due to the high dimensionality of industrial multivarlate time-series and the lack of labeled anomaly· samples, deep neural networks with the ability of learning temporal patterns in an unsupervised way have become the mainstream techniques, but there is still remaining limitations. First, they have not explicitly modeled the complex correlations between different dimensions. Second, they cannot make a balance between pattern deviation anomalies and single metric anomalies. Aiming at these limitations, this paper proposes AD-FGP, a framework for industrial multivariate timeseries anomaly detection. At)-FGP has two novel features. First, it explicitly learns the correlations between different dimensions using a graph neural network. Second, it fuses a generative model and a predictive model to detect both pattern deviation anomalies and single metric anomalies effectively. We conducted extensive experiments based on both real-world and public datasets. Experiment results show that AD-FGP has a best overall anomaly detection performance by increasing the F 1 -score 5% to 40% as compared to the baseline methods. [ABSTRACT FROM AUTHOR]