Time Series Outlier Detection: Master project Aalborg University
Author
Christensen, Mik
Term
4. term
Publication year
2021
Submitted on
2021-06-03
Pages
51
Abstract
Time series data—measurements collected over time—are important in areas such as healthcare, finance, and industry. In many real-world tasks, the unusual points matter most: for example, in an electrocardiogram (ECG), clinicians focus on irregular heartbeat patterns rather than all the normal beats. Many time series datasets are unlabeled, so we do not know in advance which observations are normal or abnormal. This makes standard classification unsuitable and creates a need for unsupervised anomaly (outlier) detection. This thesis presents six algorithms for unsupervised outlier detection in time series. They share the idea of learning what normal data look like and then flagging departures from that pattern. The approaches include isolation-based methods that separate unusual points quickly, neural networks that learn typical temporal patterns, and variational autoencoders that encode normal data and use reconstruction errors to highlight anomalies.
Tidsseriedata – data, der er målt over tid – er vigtige i områder som sundhedsvæsen, finans og industri. I praksis er det ofte de usædvanlige observationer, der betyder mest: En læge ser for eksempel efter uregelmæssige mønstre i et elektrokardiogram (EKG) frem for alle de normale hjerteslag. Mange tidsseriedatasæt mangler etiketter, der fortæller, hvad der er normalt eller unormalt. Derfor kan traditionelle klassifikationsmetoder ikke bruges, og der er behov for usuperviseret detektion af afvigelser (anomalier). Denne afhandling præsenterer seks algoritmer til usuperviseret detektion af afvigelser i tidsserier. Fælles for dem er, at de forsøger at lære, hvordan data normalt ser ud, og derefter markere punkter, der afviger. Tilgangene omfatter metoder, der kan isolere usædvanlige observationer hurtigere end normale, neurale netværk, der lærer typiske tidsmæssige mønstre, og variationale autoencodere, der indkoder normal data og bruger rekonstruktionsfejl til at finde afvigelser.
[This apstract has been rewritten with the help of AI based on the project's original abstract]