Detecting anomalies in data center physical infrastructures using statistical approaches

Abstract

Data center physical infrastructures including electrical system, cooling system, and other secondary systems support the developments of modern information technology. Early detection of the unexpected observations in physical infrastructures is of great significance to prevent the breakdown of the system and further losses. However, the state of the art technique for identifying anomalies in existing infrastructure monitoring platform mainly depends on fixed threshold method. An obvious drawback of the method is that it usually leads to a high misdetection rate. In this study, statistical anomaly detection approach is introduced to physical infrastructure monitoring. First, three important types of anomalies encountered in infrastructure monitoring platform are addressed, namely naïve point anomalies, contextual point anomalies, and level shifts. Then, a method based on Gaussian model is put forward to detect the above three anomalies. Because the proposed method can only effectively detect the naïve point anomalies; an improved approach combining the statistical test results on original and first-differenced monitoring data is provided. Performances of the proposed methods on a real data set are evaluated. Results show that the optimized anomaly detection approach has a good precision and can significantly lower the misdetection rate. In conclusion, this study will not only contribute to the improvement of existing monitoring platform but also benefit the preventive maintenance of data center physical infrastructures.