FDNN: Feature-based Deep Neural Network Model for Anomaly Detection of KPIs

Anomaly detection of KPIs (key performance indicators) has been widely applied to guarantee systems stability in real world. KPIs include response time of Web pages, CPU utilization, memory utilization, disk IO and so on. However, time series of different KPIs have different shapes, so that it is a great challenge to detect anomaly of KPIs by a simple statistical or machine learning model. In this paper, we design and implement FDNN (Feature-based Deep Neural Network) model for anomaly detection of KPIs. We present a novel feature engineering approach called MSWFeature (multiple sliding windows feature) which is more suitable to extract temporal feature for time series of KPIs. FDNN model with MSWFeature achieves good performance in F1-Score over other supervised models for anomaly detection on the studied KPIs dataset collected by the top global internet companies.