Multivariate statistical process control using dynamic ensemble methods
| dc.contributor.advisor | Weihs, Claus | |
| dc.contributor.author | Mejri, Dhouha | |
| dc.contributor.referee | Ligges, Uwe | |
| dc.date.accepted | 2015 | |
| dc.date.accessioned | 2016-01-25T11:04:57Z | |
| dc.date.available | 2016-01-25T11:04:57Z | |
| dc.date.issued | 2015 | |
| dc.description.abstract | One important challenge with some applications such as credit card fraud detection, intrusion detection and network traffic monitoring is that data arrive in streams over time and leads to changes in concepts which are known in data mining as concept drift. Thus, models analyzing such data become obsolete and efficient learning should be able to identify these changes and quickly update the system to them. The objective of this dissertation is to investigate the effectiveness of ensemble methods and Statistical Process Control (SPC) techniques in detecting changes in processes in order to improve the robustness of tracking concept drift and coping with the dynamics of online data stream processes. For reaching this objective, different heuristics were proposed. First, an improved dynamic weighted majority Winnow algorithm based on ensemble methods is proposed. Furthermore, parameters optimization based on genetic algorithm of the proposed method as well as an analysis of its robustness are investigated. Second, in order to handle the problem of concept drift while monitoring nonstationary environment using SPC tools, a time adjusting control chart based on a recursive adaptive formulas of the charting statistics is proposed. Results show that the updating charts cope much better with the nonstationarity of the environment. Also, two new heuristics are proposed based on both ensemble methods and adaptive control charts. The first is an offline learning chart model while the second is an online batch learning algorithm. Results show that quick adaptation of the system and accurate shift point identification are achieved when using both heuristics together. Also, the new adaptive ensemble charts have better performance in learning concept drifts along with a good suitability to nonlinearity and noise issues. | en |
| dc.identifier.uri | http://hdl.handle.net/2003/34465 | |
| dc.identifier.uri | http://dx.doi.org/10.17877/DE290R-16521 | |
| dc.language.iso | en | de |
| dc.subject | Ensemble methods | en |
| dc.subject | Statistical process control | en |
| dc.subject.ddc | 310 | |
| dc.title | Multivariate statistical process control using dynamic ensemble methods | en |
| dc.type | Text | de |
| dc.type.publicationtype | doctoralThesis | de |
| dcterms.accessRights | open access |