الفهرس 
INTRODUCTIONOnly 14 pages are availabe for public view
 |  | from | 128 |  |  |
| | | from | 128 | | |
Abstract
Health Management (HM) of Rolling Element Bearings (REB) continues to be of increasing importance to rotating equipment productivity, reliability, availability, and cost reduction. Fault Detection (FD) is a crucial pillar of HM as part of the evolving Prognostics and HM (PHM) philosophy enabling predictive maintenance. Finding data-driven FD Methods is an active topic of research in the PHM field. This thesis proposes an REB data-driven FD method that utilizes established machine learning/ data-driven models without needing supervised feature extraction. The method first applies principal component analysis to the REB vibration signal(s) after segmenting the signal(s) into overlapping blocks. A Hidden Markov Model (HMM) is trained on the principal components next where k-means clustering is applied for setting the HMM’s number of states. Ultimately, the trained HMM is employed, together with a Z-Score test of hypothesis, to assess the bearing health state on simulated real-time data. The method is evaluated on the REB testbed dataset provided by the center for Intelligent Maintenance Systems (IMS), University of Cincinnati, OH. Experimental results show that it outperforms the results of other methods in the literature based on the amount of data used for training the method and how early the fault was detected, while it came second in the number of tests the method succeeded in detecting the fault. Further, a Transfer Learning (TL) method is proposed and applied in conjunction with the FD method. The results show that TL from a previously trained source bearing to a new target bearing increases the generalization and robustness of the FD method as it reduces the required target bearing training data to achieve high levels of precision, recall, and accuracy.