Bearing remaining useful life estimation using an adaptive data-driven model based on health state change point identification and K-means clustering

• Published in: Measurement science & technology Vol. 31; no. 8; pp. 85601 - 85624
• Main Authors: Singh, Jaskaran, Darpe, A K, Singh, S P
• Format: Journal Article
• Language: English
• Published: IOP Publishing 01.08.2020
• Subjects: bearing health indicator; confidence value; prognostics; remaining useful life; rolling element bearings
• ISSN: 0957-0233; 1361-6501
• DOI: 10.1088/1361-6501/ab6671

Advance prediction about bearing remaining useful life (RUL) is a major activity which aims at scheduling proper future actions to avoid catastrophic events. However, the reliability of bearing life prediction models is subject to processes, such as construction of a robust bearing degradation health index, monotonicity and trendability of health index, uncertainty in construction of a failure threshold etc. Therefore, to achieve reliable bearing RUL estimates, this study proposes a fundamental framework wherein several data driven models are trained adaptively corresponding to the different bearing health states. The core idea is to selectively identify effective bearings from the training set of bearings whose failure patterns match closely with the evolving failure pattern of a bearing under operation. In each bearing, the locations of all health state change points are identified and then the training bearings are clustered into groups having similar failure trajectories using a K-means approach and developed similarity index. The proposed approach utilizes only partial data from the test bearing for RUL prediction and eliminates the need to manually pre-define a failure threshold limit. The prediction estimates are updated with every incoming data point acquired on the test bearing until failure. A cumulative function is proposed to make the trend of the adopted health indicator (HI) into being monotonic and trendable, which is then used as an input to the data driven model. A confidence value (CV) parameter is proposed to map the inputs of the data driven model, such the CV varies in a fixed range. Both simulated data and run-to-failure experimental data (IEEE PHM 2012 bearing data) have been used to demonstrate the effectiveness of the proposed method. The test results from the proposed methodology have been benchmarked with other approaches, further validating its generic character and robustness.