Anomaly Detection In The Performance Of Data Center Cooling System Devices Based On Machine Learning And Time Series Analysis
Marina Repeva, Stanislav O. Polyakov · 2025
In the context of the ongoing research into methods for the early detection of freon leakage in cooling systems within containerized data centers, the paper explores the potential of machine learning methodologies and time series analysis for the identification of anomalies in the operation of the relevant technical equipment. The state and load of this equipment are thoroughly delineated through the use of multidimensional time series. The study is grounded in data obtained by systematically collecting numerical values from multiple air conditioner sensors within a real containerized data center monitoring system. The study draws upon the distinctive characteristics of the original data, leveraging the anomaly detection capabilities of three prominent open source libraries (Merlion, Darts, and Anomaly Detection Toolkit) to organize work with diverse machine learning and time series analysis methods. These approaches are then employed to interpret the obtained anomalous values in accordance with the objectives delineated in the study. The experimental results demonstrate that the most effective methods for identifying malfunctions, such as freon leakage in data center air conditioners, approximately two and a half hours before complete equipment failure, are those based on approaches such as LevelShiftAD and VolatilityShiftAD from the Anomaly Detection Toolkit library.