Reliable Intelligent Diagnostics With Uncertainty Quantification for Mechanical System Condition Assessment

ABSTRACT Health condition assessment of mechanical systems is essential to ensure their safe and reliable operation. Although AI‐driven diagnostic techniques have advanced significantly with the development of deep learning, the reliability of such techniques is often compromised due to the lack of uncertainty quantification (UQ)—particularly the neglect of both epistemic (model‐based) and aleatoric (data‐based) uncertainties. To enhance the reliability of AI diagnostics, this study investigates the uncertainty characteristics of deep learning models by decomposing and quantifying predictive uncertainty. Specifically, four deep neural network architectures are constructed and compared for fault diagnosis using bearing vibration datasets from Southeast University and Case Western Reserve University. Vibration signals are preprocessed through overlapped sampling using the sliding window method, and time‐frequency features are extracted via continuous wavelet transform (CWT). Monte Carlo (MC) dropout is applied to estimate total predictive uncertainty (PU), which is further decomposed into aleatoric uncertainty (AU) and epistemic uncertainty (EU), enabling evaluation of model prediction confidence and identification of abnormal samples. The results demonstrate that deeper network architectures, particularly ResNet34, exhibit significantly lower epistemic uncertainty, indicating higher model confidence under complex conditions. Additionally, the use of feature‐extracted input contributes to a reduction in epistemic uncertainty, while its influence on the overall uncertainty profile is model‐dependent. These findings provide insights into the development of trustworthy intelligent fault diagnosis methods with enhanced reliability for practical health management of mechanical equipment.