TabPFN-enabled multimodal diagnostic framework with interphase interference compensation for metal oxide surge arresters

This is the authors' abstract. We don't add key points for this paper.

Reliable fault diagnosis of metal oxide surge arresters (MOSAs) requires accurate interpretation of leakage currents under interphase capacitive interference and limited fault data. This paper proposes a current–temperature multimodal diagnostic framework integrating vector-decomposition-based interphase interference compensation with a Tabular Prior-data Fitted Network (TabPFN) classifier. The proposed compensation method targets 50 Hz fundamental frequency leakage current phasors and estimates adjacent-phase capacitive currents directly from the measured three-phase current vector, thereby reducing the dependence on complete station-specific geometric modeling. A multi-physics simulation dataset covering seven operating conditions is constructed and evaluated using phase-blocked, severity-blocked, and simulation-sweep-blocked holdout strategies to enable a more independent assessment. For current-only diagnosis, numerical compensation improves the average accuracy from 0.757 to 0.802. By further incorporating temperature features, the proposed TabPFN framework achieved an average accuracy of 0.886 and an average Macro-F1 score of 0.883, which were numerically higher than those of CatBoost and LightGBM, although the differences were not statistically significant. Using in-situ leakage current monitoring data from 500 kV MOSAs manufactured by Fushun Electric Porcelain Factory, the proposed compensation method reduced the amplitude range, phase-angle range, and resistive current range by 82.3 %, 73.1 %, and 70.9 %, respectively. The proposed framework provides a promising technical foundation for intelligent monitoring and fault diagnosis of MOSAs.