A Novel Unsupervised Dead-value Detection Method for Monitoring Indicators in Data Center

For equipment monitoring, monitoring data are continuously generated, and when the monitoring data remain constant abnormally, we believe that these are dead values. Dead values are common and dead value detection is critical to subsequent data driven intelligent analysis for equipment. However, there is no specify work to study the dead value detection problem, to the best of our knowledge. In addition, due to challenges such as confusing profiles of dead values, the huge amount of monitoring data, and nonstationarity, existing anomaly detection methods are invalid. In this paper, we propose an effective dead value detection method consisting of two steps: dead value scoring and dead value detecting. In our evaluation, we analyze the monitoring data of equipment in the data center, and summarize six representative monitoring indicators with dead values as dataset. The evaluation experiments indicate the proposed dead value detection method achieves an average F1 score of 0.93, significantly outperforming the best performing baseline detection approaches by 117.5% on average.