Complex-impedance features for transformer winding-fault diagnosis: Finite-element classification with laboratory frequency-response corroboration

Automated interpretation of transformer frequency-response analysis (FRA) often retains impedance magnitude but discards signed real and imaginary components. This study evaluates complex ( R, X ) features for five-class winding-fault classification and fault-specific severity estimation. A finite-element model supplies network parameters for 210 labelled cases, and a state-space solution supplies the complex frequency response. Evaluation uses repeated stratified five-fold cross-validation (20 repeats) with one common healthy reference and fold-local preprocessing. Relative to magnitude-only features, complex frequency-domain features increase LightGBM accuracy from 93.57 % to 94.95 % ( + 1.38 percentage points, paired p = 0.0012 ). Adding descriptors of the inverse-transformed spectrum changes accuracy to 95.10 % ( + 0.14 points, p = 0.356 ), confirming that these descriptors are only a re-encoding of the same spectrum. A stacked ensemble reaches 95.45 % accuracy and 94.87 % macro-F1, compared with 95.93 % and 95.52 % for XGBoost; the difference is unresolved ( p = 0.169 ), so no superiority is claimed. Laboratory impedance sweeps from one disc-winding specimen corroborate the directions of inductance and anti-resonance changes for controlled axial, disc-space, radial and short-circuit interventions. They also reveal a low-frequency resistance-resolution floor and the absence of decay structure in the inverse-transformed response. A public converter-transformer magnitude–phase record additionally confirms that signed R and X inputs can be reconstructed from an in-service measurement. Because the measured set contains only one healthy sweep and ten retained deformations, it tests physical premises but does not externally validate the classifier. The reported predictive performance therefore remains specific to the simulated dataset.