Time series anomaly detection is an important task applied in various fields such as finance, healthcare, and meteorological monitoring. Information about the locations of anomalies in various time series is rarely available in real-life settings. Therefore it is necessary to use unsupervised methods for anomaly detection.
In this thesis, we compared four unsupervised anomaly detection methods: matrix profile, LOF, isolation forest, and autoencoder. The selected methods belong to different fields of computer science and statistics such as outlier detection, machine learning, and deep learning. They also differ in the way they detect anomalies: measuring distances between time series, using tree structures, and reconstruction.
To evaluate the methods, we used the UCR Anomaly Archive dataset and an artificially generated set of time series with anomalies. We also applied a dynamic sliding window size to all methods, which was determined according to the characteristics of different time series. We also used the selected methods to detect anomalies detection on real production data provided by the company Senso4s Ltd..
The best results were obtained with the matrix profile method, followed by the LOF method with slightly worse results and a shorter running time.
|
