Time Series Analysis via Shapelet Generation
| Type of publication: | Mastersthesis |
| Citation: | |
| Type: | Master Thesis |
| Year: | 2018 |
| Month: | January |
| School: | TU Darmstadt |
| Abstract: | Nowadays all kinds of sensors track massive amounts of data every second. Capturing and analyzing this data is becoming more and more interesting for industries concerned with maintaining machinery. Companies want to predict the health status of equipment to reduce maintenance costs. The prediction is usually based on time series of sensor data. In recent times shapelets as feature become more and more popular in research. In this thesis an algorithm for finding outliers in time series is investigated. Unlike classic shapelet finder, the method examined uses shapelet generation and pseudo classes to handle unlabeled data. This work is concerned with the ability to generate shapelets and the classification performance based on it. The results show that the performance introduced by the authors cannot be reproduced in this work. A major problem is the generation of good matching shapelets in data sets with more than one class. Although generation of shapelets on unlabeled data is a promising approach to deal with the vast amounts of recorded time series in industrial context. In its current state, the method examined in this work is not able to perform the given task. |
| Keywords: | |
| Authors | |
|
Topics
|
|
|
|
