Graph embedding-based Anomaly localization for HVAC system

Abstract

As a major energy consumption system in buildings, anomaly detection on multivariate time series monitored by sensors in HVAC systems has been a significant challenge. However, existing data-driven anomaly detection methods are dedicated to improving the performance of abnormal time interval detection that fails to localize abnormal time series within abnormal time intervals. In this study, we design a novel anomaly localization method based on graph embedding and apply it to the HVAC system. The proposed method aims to learn feature representations of time intervals and time series by graph embedding and to detect abnormal time intervals and time series by identifying outlier degrees of embedding vectors. It constructs dynamic graphs representing time intervals by extracting correlations between multivariate time series and proposes an entropy-based graph embedding model to embed the dynamic graph. In the embedding space, the local outlier factor is used to measure abnormal embedding vectors to detect abnormal time intervals. After detecting abnormal time intervals, large-scale information network embedding is applied to embed each time series within the abnormal time interval for detecting abnormal time series. Our method is evaluated on three different abnormal patterns of the HVAC system. The experimental results reveal that the F1 scores of our method beat the existing methods in terms of abnormal time interval detection and abnormal time series detection of the HVAC system. © 2023