Analysis of Transformer Health Index Using Bayesian Statistical Models

Health index (HI) is a very useful tool for representing the overall health of a complex asset, such as the power transformer, due to the fact that it quantifies equipment condition based on different criteria that are related to the longterm degradation factors that cumulatively lead to the asset's end-of-life. The main concern with HI computation is with the practical management of the numerous criteria that are combined in different ways (with proprietary information and associated weighting factors) to produce a HI value. Hence, several authors have proposed different approaches to the HI calculation, e.g., analytical expressions, logistic regression, fuzzy logic, support vector machines, and artificial neural networks. This paper proposes using Bayesian multinomial logistic regression for the HI calculation. This approach offers high flexibility with multiple metric and/or nominal predictors, including correlation and interaction between predictors, and acknowledges the fact that the transformer HI is described with three to five categories. It further offers high model interpretability and benefits from the Bayesian ability to quantize uncertainty in model parameters.