Open AccessAn Enhanced Seasonal-Hybrid ESD Technique for Robust Anomaly Detection on Time Series
Author(s) -
Rafael Vieira,
Marcos A. Leone Filho,
Robinson Semolini
Publication year - 2018
Language(s) - English
Resource type - Conference proceedings
DOI - 10.5753/sbrc.2018.2422
Subject(s) - anomaly detection , benchmark (surveying) , computer science , outlier , context (archaeology) , anomaly (physics) , studentized range , series (stratigraphy) , data mining , time series , range (aeronautics) , process (computing) , artificial intelligence , machine learning , engineering , mathematics , statistics , paleontology , physics , geodesy , condensed matter physics , standard deviation , biology , aerospace engineering , geography , operating system
Nowadays, time series data underlies countless research activities. Despite the wide range of techniques to capture and process all this information, issues such as analyzing large amounts of data and detecting unusual behaviors on them still pose a great challenge. In this context, this paper suggests SHESD+, a statistical technique that combines the Extreme Studentized Deviate (ESD) test and a decomposition procedure based on Loess to detect anomalies on time series data. The proposed technique employs robust metrics to identify anomalies in a more proper and accurate manner, even in the presence of trend and seasonal spikes. Simulation studies are carried out to evaluate the effectiveness of the SH-ESD+ using the published Numenta Anomaly Benchmark (NAB) collection. Computational results show that the SH-ESD+ performs consistently when compared against state-of-the-art and classic detection techniques.