Learning Hierarchical Spatial-Temporal Graph Representations for Robust Multivariate Industrial Anomaly Detection

Multivariate time series anomaly detection is one of the most indispensable yet troublesome links in complex industrial processes. The main challenge lies in discovering the representative patterns for collective or contextual anomalies among interconnected sensory data streams, which has been largely hampered by inefficient spatial-temporal feature extraction and suboptimal decision criteria under the scarcity of positive training samples. This article goes beyond the common limitations of the existing methods, and novelly proposes Hierarchical Spatial-Temporal grAph Representation (HiSTAR). It processes the data with strong structural inductive biases through latent spatial-temporal graph modeling, yet requiring no predefined topological priors for the sensor network. A discriminative decision boundary is constructed by learning hierarchical normality-enclosing hyperspheres on the produced graph-structure representations. In this way, HiSTAR not only presents superior anomaly detection performance, but also provides consistent anomaly localization results. The efficacy of the proposed method is experimentally corroborated through three industrial case studies.

I. Introduction

Anomaly detection aims to identify the observations or events that deviate from the expected behaviors of the majority [1]. It plays a fundamental rule in most industrial applications, so as to provide critical health status information and support preventive maintenance for the running systems. Nowadays, with the rapid development of AI technologies in industrial fields, it has been an irresistible trend to discover system anomalies in a data-driven manner. Although plenty of algorithms have been performed on time series anomaly detection, most of them are designed for the univariate case. The ever-growing sensing and computing capabilities have revealed their weakness in scenarios, where large-scale sensory data can be collected from different locations of the integrated monitoring systems. It is necessary to develop robust anomaly detectors for multivariate time series data.