Machine Learning Assisted Fuzzy Logic for HVDC Converter Transformers Health Monitoring

Abstract

High-voltage direct current (HVDC) converter transformers are among the costliest and most critical assets in a transmission network. When one fails without warning, the result can be a wide-area outage and heavy financial loss. Routine condition monitoring still relies largely on fixed limits and the judgement of experienced engineers, which makes it difficult to capture the nonlinear and uncertain way insulation actually degrades. We propose a hybrid framework in which machine learning supports fuzzy logic for health monitoring of these transformers. Dissolved Gas Analysis (DGA) concentrations, winding and oil temperature, moisture content, load current, partial discharge activity, and power factor are first passed to supervised classifiers, whose task is to expose latent fault patterns. The resulting fault probabilities and derived features are then supplied to a fuzzy inference system. It fuzzifies the inputs, applies an expert-derived rule base, and defuzzifies the result into a continuous Transformer Health Index (THI) that can be computed in real time at an HVDC substation. Two extensions are implemented and evaluated quantitatively here: MICE-based imputation of missing data and a wider set of oil, dissolved-gas, thermal, and operational features. A third extension, an adaptive fuzzy-ML fusion layer that weights each output by its confidence and by data quality, is still being integrated, and its initial validation is under way.