A framework towards quantifying aleatoric and epistemic uncertainties in rolling element bearing condition monitoring experiments

Abstract

Rolling element bearings (REBs) are critical components in rotating machinery, and their condition monitoring (CM) is essential for predictive maintenance. However, experimental data used for developing and validating CM techniques often contain uncertainties arising from sensor noise, environmental conditions, and test bench disassembly and reassembly. These experimental datasets’ uncertainties, categorised as epistemic and aleatoric, are typically overlooked in developing and validating these CM techniques.

This study developed a framework to quantify epistemic and aleatoric uncertainties in REB CM datasets. An experimental methodology was developed to introduce and quantify these uncertainties. The experimental methodology was applied to a bearing test bench at the University of Pretoria. Four cases were designed to quantify epistemic uncertainty introduced by the bearing assemblymodification: (1) baseline configuration (without any modification), (2) bearing preload removal and reinstallation, (3) bearing preload and partial bearing removal (all components except for the outer race), followed by reinstallation of the components, and (4) removal of the preload with the complete bearing whereafter all the components were reinstalled. Aleatoric uncertainty was captured by taking five consecutive measurements per case at fixed intervals under constant speed. The collected signals were further divided into full and segmented windows for comparative analysis and were analysed using the developed envelope analysis-based methodology.

Before detailed uncertainty analysis commenced, the influence of key envelope analysis parameters on the results’ interpretation was investigated using a simulated and an experimental signal. These parameters were investigated so that their effect on the experimental uncertainties is minimised.

Unlike most studies in the literature, which focus on model-level uncertainties, this work emphasises data-level uncertainties arising from the experimental setup and data collection process. By quantifying and analysing how these uncertainties propagate through feature extraction, the key findings demonstrate that aleatoric uncertainties induce measurable variability between repeatedmeasurements. In contrast, epistemic uncertainty causes significant shifts between cases. This shift is attributable to the changes made to the bearing in the bearing housing, thereby creating out-of-distribution effects on diagnostic features. This shows the importance of accounting for both inherent data variability and experimental setup changes to enhance the trustworthiness and robustness of REBs fault diagnosis technique.

This work presents a novel perspective in REB diagnostics, underscoring the importance of data variability and quality. The experimental and analysis framework established here provides the groundwork for integrating experimental dataset uncertainty-awareness into artificial intelligence (AI) based CM techniques (e.g., machine learning).