Gaussian Process Regression for an Accurate and Uncertainty-Aware Winding Insulation Degradation Prediction

Accurate prediction of winding insulation degradation path is critical for preventing catastrophic equipment failures and optimizing maintenance schedules in electric motors (EMs). Existing methods, such as those based on monitoring high-frequency electrical parameters, often rely on point estimates and neglecting the inherent uncertainties associated with real-world degradation processes. This paper proposes a novel approach utilizing Gaussian Process Regression (GPR) to address this limitation. Building upon recent advancements in high-frequency electrical parameter monitoring in which identifying inter-turn insulation creep is a key degradation indicator, this work adopts GPR to predict the degradation path. GPR offers a powerful framework for incorporating uncertainty quantification into the prediction process. It not only excels at interpolation within the observed data range but also provides a distribution of possible future degradation values. This probabilistic approach acknowledges the variability present in both real-world measurements and the inherent process variability of insulation degradation. The prediction results from proposed GPR-based approach are compared to a nonlinear Wiener-processbased model as a conventional method, and a state-of-the-art optimization algorithm. The estimation accuracy in the worst case scenario of the proposed method gives an error of 0.7% which is more accurate than 4.2%, and 50% resulted from the Wiener-process-based model and the commercial optimization solver respectively. These results demonstrate a significant improvement in estimation accuracy by effectively handling both data and process-related uncertainties.