Rolling element bearings (REBs) are one of the most critical mechanical components. Their failures can lead for catastrophic failures which might include great loss in economy or even in the lives of people. REBs are inherently dynamic and they demonstrate complex vibration behaviour where conventional vibration -based fault diagnosis methods might not give sensitive indicators of the presence of the defects. This thesis investigates the singular spectrum analysis (SSA) capabilities as completely data-based fault diagnosis method in REBs. The SSA is used to decompose the bearing vibration acceleration signals in a certain number of principal components having the trend, periodical components and structure-less noise. This thesis develops two methodologies to use SSA in different ways and for different purposes. The first methodology uses the SSA (i.e only the decomposition stage) to create a baseline space from healthy bearing vibration signals. Then, any new signals are projected onto this baseline space. From these projections, features are made and used for fault diagnosis purposes. In the second methodology, the SSA contributes to the development of an advanced signal pretreatment that efficiently improves representing the nonstationary bearing vibration signals by linear time invariant autoregressive (LTIVAR) model. Then the coefficients of LTIVAR model are used as features for fault diagnosis purposes.The two methodologies have been validated by using experimental data obtained from three different bearing test rigs. The data used in the analysis covers different defect locations and different defect severities. The results of both methodologies, in terms of correct classification, were compared to some other recent methodologies. In comparison, it is shown that both methodologies have a very good performance and they are superior to those methodologies.The thesis offers simple and efficient methodologies for a complete fault diagnosis in terms of fault detection, identification and severity estimation. Thus, these methodologies have a potential possibility for automation of the entire process of each method.
|Date of Award||24 Oct 2016|
- University Of Strathclyde
|Supervisor||Irina Trendafilova (Supervisor) & Gareth Pierce (Supervisor)|