Uncertainty-aware fault mode diagnostics for condition-based maintenance of bearing

Condition-based maintenance (CBM) relies on diagnostic information inferred from condition monitoring data to support maintenance decisions. While diagnostics are inherently uncertain in practice, most CBM frameworks either assume perfectly observable degradation states or implicitly assume fully reliable diagnostics. This limitation becomes more critical in systems with multiple fault modes, where diagnostic ambiguity is unavoidable and may strongly affect maintenance performance. This paper studies CBM of bearings under multiple fault modes. First, a physics-based kinematic model is developed to simulate vibration signals associated with outer race, inner race, and ball defects, extending an existing bearing model to include ball defects. Based on the simulated data, a datadriven fault diagnostics model is constructed using a 1D convolutional neural network with Monte Carlo dropout to estimate the posterior probability of each fault mode. The estimated diagnostic posterior is then integrated into maintenance decision-making through threshold-based condition-based replacement policies. Two policies are proposed: one triggers preventive replacement only when diagnostic confidence is sufficiently high, and the other triggers inspection when diagnostic uncertainty is high. A numerical case study demonstrates that relying on point-estimate diagnostics increases maintenance costs, whereas explicitly accounting for diagnostic uncertainty substantially reduces costs. The results highlight the importance of uncertainty-aware diagnostics for cost-effective condition-based maintenance, particularly when inspections are costly and diagnostics are imperfect.

Keywords: condition-based maintenance, fault diagnostics, diagnostic uncertainty, Monte Carlo dropout, bearing fault simulation.