Extended kernel density estimation for anomaly detection in streaming data

Abstract

Machine Learning algorithms based on kernel density estimation (KDE) are said to be well suited for anomaly detection. However, existing approaches mainly cover point anomaly detection. Many industrial applications also require detecting drifts, contextual and collective anomalies. Due to the demand for small latencies, edge computing gains significance and hybrid models are no adequate solution. The main contribution is the EEM-KDE algorithm, which includes extensions of the KDE for detecting all aforementioned types of anomalies in streaming data in the edge. Finally, the proposed algorithm is evaluated showing the capability to detect different types of anomalies using an industrial control unit.