Anomaly Detection in Data Center Infrastructure Using Graph Neural Networks on Multivariate Sensor Streams

Abstract

Early detection of infrastructure anomalies in data centers — such as impending hardwarefailures, cooling system degradation, and network bottlenecks — is critical for maintainingservice reliability and avoiding costly unplanned downtime. Traditional threshold-based monitoring and univariate statistical methods fail to capture the complex spatial and temporaldependencies among thousands of interrelated sensors in modern facilities.