Anomaly detection of timeseries: A comparison of statistical vs classical machine learning vs deep learning approaches
| Type of publication: | Mastersthesis |
| Citation: | |
| Year: | 2019 |
| Month: | November |
| School: | TU Darmstadt |
| Abstract: | Anomaly detection on timeseries data has been an important research field for a long time. While the original anomaly detection methods have been based on statistical approaches, in recent years more and more machine learning algorithms have been developed to detect anomalies on time series. Furthermore, many researchers tried to improve these techniques using neural networks. In the light of the continuously increasing number of anomaly detection methods of these three classes, the research community suffers from the lack of a broad comparative evaluation of statistical, machine learning and deep learning methods. This thesis tries to overcome this shortcoming by studying 20 univariate and 24 multivariate anomaly detection methods from the mentioned categories. Additionally, the evaluation will be done on publicly available datasets, which serve as benchmarks for time series anomaly detection. To provide a reliable comparison, in addition to the accuracy of each method, the computation time of the algorithms is measured. By analysing univariate as well as multivariate timeseries, we hope to provide a thorough insight about the performance of these three classes of anomaly detection approaches. |
| Keywords: | |
| Authors | |
|
Attachments
|
|
|
|
|
Topics
|
|
|
|
