Uncertainty Quantification in Fault and Degradation Analysis of Rolling Element Bearings

Abstract

This study presents an innovative framework for deriving degradation models of rolling element bearings through uncertainty quantification. Natural open-source run-to-failure experimental datasets, XJTU-SY and PRONOSTIA, were used in this research to investigate the uncertainty at the incipient fault point (IF) and the end-of-life point (EOL) among identical ball bearings under the same operational conditions. This study answers the research question: Can data-driven analysis based on entropy and uncertainty quantification enhance explainability in degradation model determination? The objectives of this paper are to (1) identify the unknown fault types, (2) quantify the uncertainty of the IFs and EOLs, and (3) determine the degradation model considering uncertainty. Fault diagnosis was achieved using a wavelet entropy-based approach integrated with power spectral analysis and clustering via K-means to identify and classify fault types probabilistically. Sensitivity analysis and feature selection were applied in a recursive method to reduce the dimensionality, enhancing model accuracy to 90%. Fault diagnosis contributes to quantifying the uncertainty of the IF and EOL for similar fault-induced bearings using the maximum entropy (MaxEnt) principle. This translated these critical parameters into deterministic descriptors with probability rather than fixed deterministic values. Due to limited data from both datasets, the study employs MaxEnt again to define probability density functions used to generate the degradation model. The results demonstrated that the probabilistic degradation model effectively captures the inherent variability in degradation processes. This methodology is extendable to other engineering systems, offering a versatile tool for predictive maintenance and remaining useful life estimation.