Uncertainty quantification of faults in rotating machines

Rotating machines such as bearings, turbines, motors, gearboxes, and compressors are indispensable in modern industry, forming the mechanical backbone of key sectors including energy, manufacturing, oil and gas, and water treatment. Automation and digitalization have increased the demand for effective predictive maintenance to ensure peak performance, operational availability, and long-term reliability of rotating equipment. Thus, the goal of this dissertation is to enhance predictive maintenance of rotating machines. This study answers the question, “Can uncertainty quantification enhance the fault prediction in rotating machines?” Three objectives were constructed to facilitate the answer. The first objective was to construct probabilistic degradation models of ball bearings, given their immense importance in every rotating machine, because probabilistic degradation models adopted in literature do not consider incipient fault point deduction and assume predefined probability distributions for their modeling. Using the maximum entropy principle, a probabilistic degradation model was proposed to incorporate the uncertainty surrounding the incipient fault point and the end-of-life timestamp of a bearing. The study further proposed an incipient fault deduction approach that accurately captured early faults, essential for an effective prediction model. The second objective was to evaluate the epistemic uncertainty in the diagnostics of a wind turbine’s planetary gearbox. Due to the increasing demand for renewable energy, the power rating for wind turbines has grown significantly. The increase in power accounts for a corresponding rise in uneven loading distributions in the gearbox. A probabilistic fault-detection framework was proposed for a planetary gearbox with an induced fault over increasing power ratings, utilizing a Bayesian model to measure the epistemic uncertainty in relation to the accuracy of the prediction. The third objective was to calculate the epistemic and aleatory uncertainties in prognostics of turbofan engines. Considering their large and complex nature, the variation in usage patterns contribute to uncertainty in the degradation process, and less relevant fault data could further cause inaccuracies and representation errors in label predictions. This study finally proposed an uncertainty-aware approach for estimating the remaining useful life of turbofan engines through a Bayesian recurrent variational encoding network, including uncertainty calibration and disentanglement, which outperformed many existing methods in the literature after redundancy issues were resolved. Thus, predictive maintenance strategies for fault prediction are enhanced through calculated degradation models, load distribution-based fault diagnostics, and uncertainty-aware prognostics under varying conditions in rotating machines of various complexities.