G5 Doctoral dissertation (article)
Data-driven methods for diagnostics of rolling element bearings (2019)
Kansanaho, J. (2019). Data-driven methods for diagnostics of rolling element bearings [Doctoral dissertation]. Jyväskylän yliopisto. JYU dissertations, 151. http://urn.fi/URN:ISBN:978-951-39-7936-2
JYU authors or editors
Publication details
All authors or editors: Kansanaho, Jarno
eISBN: 978-951-39-7936-2
Journal or series: JYU dissertations
eISSN: 2489-9003
Publication year: 2019
Number in series: 151
Number of pages in the book: 1 verkkoaineisto (75 sivua, 39 sivua useina numerointijaksoina, 17 numeroimatonta sivua)
Publisher: Jyväskylän yliopisto
Place of Publication: Jyväskylä
Publication country: Finland
Publication language: English
Persistent website address: http://urn.fi/URN:ISBN:978-951-39-7936-2
Publication open access: Openly available
Publication channel open access: Open Access channel
Abstract
This thesis focuses on the research and development of the data-driven methods used to diagnose rolling element bearings (REBs) and evaluates the software architectural design of these data-driven methods. REBs are vulnerable components in machinery. Vibration-based condition monitoring is a very popular methodology for monitoring the health of REBs. This research started with the development of methods to analyze and detect incipient local faults of REBs using vibration measurements. The main goal was to find weak vibration signatures generated by local faults in REBs. As a result, a flexible simulator was developed to analyze the vibrations of bearing faults and to evaluate vibration analysis methods, and a spline wavelet-based algorithm were introduced for fault detection. An incipient bearing fault will become enlarged if a machine is run and the faulty bearing has not been replaced. The identification of different lifetime stages of wear evolution is part of the input data for bearing diagnostics and prognostics. A method to detect different lifetime stages of REBs according to their vibration signals was proposed based on an unsupervised learning method. The result of the unsupervised method was exploited in early fault detection utilizing supervised methods. It is important to estimate the severity of a fault, and size is probably the best proxy for severity. Estimating the fault size of defective REBs is one of the top challenges in bearing diagnostics, especially when vibration measurements are used to determine the state of health. A novel method for feature ranking to estimate fault sizes for REBs was presented. Black-box classifiers were applied to detect non-linear relations between features, and it was concluded that the best metrics for basic diagnostics are not necessarily the best qualities for fault size estimation. The final part of this research focuses on design at system-level. Software framework designs encapsulate fault detection and remaining useful life (RUL) estimation methods. As part of the tribotronic system, the object-oriented framework considers bearing applications and potentially extends them to other mechanical applications.
Keywords: bearings; condition monitoring; defects; oscillations; signal analysis; algorithms; machine learning; application frameworks; tribology
Free keywords: rolling element bearing; bearing diagnostics; vibration analysis; feature extraction; machine learning; tribological system; software framework
Contributing organizations
Ministry reporting: Yes
Reporting Year: 2019