Uncertainty Management in the Prediction of the Remaining Useful Life of Ball Bearings and Li-ion Batteries

Effective uncertainty management is critical in predicting the remaining useful life (RUL) of ball bearings and Li-ion batteries. Accurate RUL prediction enables proactive maintenance strategies, reducing downtime and optimizing resource allocation. The goal of this research is to enhance the reliability and extend the operational life of engineering systems. This study's research question is: Can RUL predictions be enhanced through uncertainty management? The research question is answered through three specific aims. The first specific aim is to enhance the calculation of the incipient fault point (IF) of ball bearings and Li-ion batteries. To achieve the first specific aim, a methodology was developed to calculate the IF based on the Multiscale Diversity Entropy approach. This approach enables the proposed methodology to capture the physics of the problem and better predict the IF. The methodology is applied to bearing and battery datasets, and the results showed that the developed methodology better calculates the IF when compared to the standard method of calculating IF. The second objective is to determine the epistemic and aleatoric uncertainty in the prediction of RUL, considering multiple failure modes of ball bearings and Li-ion batteries. The second objective was achieved through uncertainty calculation on the predicted RUL using the developed IF method and Bayesian Neural Network. This approach was applied to the same bearing (multiple failure modes are known) and the battery datasets (multiple failure modes are unknown); from this approach, several insights can be extracted into the relationship between failure modes and uncertainty. Engineering systems with the same failure mode tended to have similar aleatoric and epistemic uncertainty signatures, indicating that each failure mode has a distinct fingerprint. In contrast, systems with multiple failure modes have more uncertainty sources, leading to considerable total uncertainty (aleatoric and epistemic uncertainty). So, critical failure tended to have more uncertainty. Finally, the epistemic uncertainty in a neural network (NN) was reduced to enhance the prediction of the RUL of ball bearings. A sensitivity analysis using the Hilbert-Schmidt Independence Criterion was performed on the uncertainty quantification model to improve the model's accuracy and reduce epistemic uncertainty. The NN with the Bayesian last layer was the approach that performed best under the proposed nonlinear conditions. The sensitivity analysis showed which hyperparameters have more effect on the NN performance. The sensitivity analysis was then used in a Two-Step Gaussian Process-based Bayesian Optimization to minimize the epistemic uncertainty and maximize the accuracy of RUL prediction. The proposed uncertainty management frameworks and methodologies contribute to enhancing the reliability and extending the operational life of engineering systems by offering actionable insights for early fault detection, accurate remaining useful life prediction under multiple failure modes, and robust neural network-based prognostics, thereby improving maintenance strategies, optimizing resource allocation, and enabling the design of more resilient systems in the face of both aleatoric and epistemic uncertainties.