Time series modeling for anomaly detection is a challenging task. In this presentation we describe the approach followed while studying various sensor time series of the INFN CNAF data center related to chiller, ups and electrical plants.
We show how the variables are correlated and the anomalies identified with the application of traditional statistical and machine learning techniques (such as DBSCAN). The anomalies discovered with the various techniques are comparable for some sensors.
Then we show how it is possible to create a Graph Neural Network to perform the same type of identification and we evaluate its outputs. We perform ablation studies to understand if the complexity of this task, over this particular dataset, needs such a complex architecture or it is possible to achieve the same results with simpler networks.
