Machine Learning-Based Anomaly Detection in Public Sector Data Center Monitoring

Machine Learning-Based Anomaly Detection in Public Sector Data Center Monitoring

Yashar Abbasloo Soumeeh1

1) University Lecturer, Senior Data Center Engineer, Karaj Fire Department and Safety Services Organization

IOI Code : ACCD-ACAZFD
Publication : 10th International Conference on Applied Researches in Science & Engineering (carseconf.com)
Abstract :
    Abstract Operating a public-sector data center means watching thousands of metrics from servers, networks and applications. Fixed alarm thresholds are easy to understand but cannot follow daily and weekly load cycles, so they miss problems that occur at quiet times and cannot detect combinations of metrics that are individually normal. This paper examines machine-learning-based anomaly detection for IT infrastructure monitoring as part of the broader practice of artificial intelligence for IT operations. We compare five detectors on synthetic multivariate telemetry with 20 injected anomalies per run, over ten independent runs: fixed thresholds, rolling z-scores, a seasonal-baseline Mahalanobis detector and two Isolation Forest variants. Fixed thresholds detect 36% of the events, whereas the seasonal Mahalanobis detector detects all of them with a point-wise F1 score of 0.94, at the price of about 1.3 false alarm episodes per day; requiring two consecutive alarms removes the false alarms completely and delays detection by about 13 min. Default Isolation Forests detect 78% and 89% of the events. The results are obtained on synthetic data and illustrate mechanisms, not field performance. The paper also discusses evaluation practice, alert fatigue, data protection and the governance of AI systems in public organizations.

Keywords : anomaly detection aiops machine learning it monitoring