Consistent Anomaly Detection and Localization of Multivariate Time Series via Cross-Correlation Graph-Based Encoder–Decoder GAN

Multivariate time series is widely derived from industrial facilities, such as power plants, manufacturing machines, spacecraft, digital devices, and so on, and anomaly detection and location is of great importance to industrial preventive maintenance. However, anomalies in multivariate time series always result from their unusual change of temporal or correlative features, and it is challenging to capture these complex characteristics. Besides, achieving consistent anomaly detection and location performance is also a tricky issue. In this article, a novel anomaly detection and location framework that combines generative adversarial networks and autoencoder is proposed to capture time dependent and correlation features of multivariate time series with the need of anomalous sequences. First, multitime scale correlation computation is utilized to encode multivariate time series into multiple cross correlation graphs, which can be fed into the proposed deep architecture for extracting more distinguishable features. On this basis, a robust cost function with multiple loss issues is designed, and reconstruction matrix deviation from original space of encoder–encoder structure is utilized to detect and locate abnormal time series, ensuring the consistency of detection and location tasks and the framework reliability. Extensive experiments on five industrial datasets are conducted to indicate our model is a generic and excellent framework for anomaly detection and location of multivariate time series.