Use this URL to cite or link to this record in EThOS:
Title: Features extraction from time series
Author: Yao, Xinxin
ISNI:       0000 0004 8504 7683
Awarding Body: University of Sheffield
Current Institution: University of Sheffield
Date of Award: 2019
Availability of Full Text:
Access from EThOS:
Access from Institution:
Time series can be found in various domains like medicine, engineering, and finance. Generally speaking, a time series is a sequence of data that represents recorded values of a phenomenon over time. This thesis studies time series mining, including transformation and distance measure, anomaly or anomalies detection, clustering and remaining useful life estimation. In the course of the first mining task (transformation and distance measure), in order to increase the accuracy of distance measure between transformed series (symbolic series), we introduce a novel calculation of distance between symbols. By integrating this newly defined method to symbolic aggregate approximation and its extensions, the experimental results show this proposed method is promising. During the process of the second mining task (anomaly or anomalies detection), for the purpose of improving the accuracy of anomaly or anomalies detection, we propose a distance measure method and an anomalies detection calculation. These proposed methods, together with previous published anomaly detection methods, are applied to real ECG data selected from MIT-BIH database. The experimental results show that our proposed outperforms other methods. During the course of the third mining task (clustering), we present an automatic clustering method, called AT-means, which can automatically carry out clustering for a given time series dataset: from the calculation of global average time series to the setting of initial centres and the determination of the number of clusters. The performance of the proposed method was tested on 10 benchmark time series datasets obtained from UCR database. For comparison, the K-means method with three different conditions are also applied to the same datasets. The experimental results show the proposed method outperforms the compared K-means approaches. During the process of the fourth mining task (remaining useful life estimation), all the original data are transformed into low-dimensional space through principal components analysis. We then proposed a novel multidimensional time series distance measure method, called as multivariate time series warping distance (MTWD), for remaining useful life estimation. This whole process is tested on the CMAPSS (Commercial Modular Aero Propulsion System Simulation) datasets and the performance is compared with two existing methods. The experimental results show that the estimated remaining useful life (RUL) values are closer to real RUL values when compared with the comparison methods. Our work contributes to the time series mining by introducing novel approaches to distance measure, anomalies detection, clustering and RUL estimation. We furthermore apply our proposed methods and related methods to benchmark datasets. The experimental results show that our methods are better than previously published methods in terms of accuracy and efficiency.
Supervisor: Wei, Hua-Liang Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available