Access Type
Open Access Dissertation
Date of Award
January 2017
Degree Type
Dissertation
Degree Name
Ph.D.
Department
Industrial and Manufacturing Engineering
First Advisor
Ratna B. Chinnam
Second Advisor
Dimitar Filev
Abstract
Applications of Condition-Based Maintenance (CBM) technology requires effective yet generic data driven methods capable of carrying out diagnostics and prognostics tasks without detailed domain knowledge and human intervention. Improved system availability, operational safety, and enhanced logistics and supply chain performance could be achieved, with the widespread deployment of CBM, at a lower cost level. This dissertation focuses on the development of a Mutual Information based Recursive Gustafson-Kessel-Like (MIRGKL) clustering algorithm which operates recursively to identify underlying model structure and parameters from stream type data. Inspired by the Evolving Gustafson-Kessel-like Clustering (eGKL) algorithm, we applied the notion of mutual information to the well-known Mahalanobis distance as the governing similarity measure throughout. This is also a special case of the Kullback-Leibler (KL) Divergence where between-cluster shape information (governed by the determinant and trace of the covariance matrix) is omitted and is only applicable in the case of normally distributed data. In the cluster assignment and consolidation process, we proposed the use of the Chi-square statistic with the provision of having different probability thresholds. Due to the symmetry and boundedness property brought in by the mutual information formulation, we have shown with real-world data that the algorithm’s performance becomes less sensitive to the same range of probability thresholds which makes system tuning a simpler task in practice. As a result, improvement demonstrated by the proposed algorithm has implications in improving generic data driven methods for diagnostics, prognostics, generic function approximations and knowledge extractions for stream type of data.
The work in this dissertation demonstrates MIRGKL’s effectiveness in clustering and knowledge representation and shows promising results in diagnostics and prognostics applications.
Recommended Citation
Tseng, Fling Finn, "Evolving Clustering Algorithms And Their Application For Condition Monitoring, Diagnostics, & Prognostics" (2017). Wayne State University Dissertations. 1750.
https://digitalcommons.wayne.edu/oa_dissertations/1750