SLMAD: Statistical Learning-Based Metric Anomaly Detection

Technology companies have become increasingly data-driven, collecting and monitoring a growing list of metrics, such as response time, throughput, page views, and user engagement. With hundreds of metrics in a production environment, an automated approach is needed to detect anomalies and alert potential incidents in real-time. In this paper, we develop a time series anomaly detection framework called Statistical Learning-Based Metric Anomaly Detection (SLMAD) that allows for the detection of anomalies from key performance indicators (KPIs) in streaming time series data. We demonstrate the integrated workflow and algorithms of our anomaly detection framework, which is designed to be accurate, efficient, unsupervised, online, robust, and generalisable. Our approach consists of a three-stage pipeline including analysis of time series, dynamic grouping, and model training and evaluation. The experimental results show that the SLMAD can accurately detect anomalies on a number of benchmark data sets and Huawei production data while maintaining efficient use of resources.