Physics‐Informed Multimodal Learning for Transformer‐Level Assessment of Nanocrystalline and Commercial Cores

ABSTRACT

This study develops a Multimodal Enhanced Physics‐Informed Learning (MEPI) framework for transformer‐level assessment using experimental magnetic and electrical measurements. Two toroidal prototypes, one incorporating an iron‐based nanocrystalline core and the other a commercial reference core, were evaluated. Of the 900 acquired measurements, 862 passed quality control, and 90 were reserved for final testing. MEPI combines a processed flux‐density waveform with nine predictive features to jointly estimate transformer efficiency, residual loss, and an Arrhenius‐informed relative thermal‐stress proxy (LSP), while excluding measured core temperature from the predictive inputs. On the final test set, MEPI achieved R ² values of 0.9818, 0.9706, and 0.9163, with corresponding MAEs of 1.1457 percentage points, 0.01369 W, and 0.02795. Matched downstream analyses showed the most pronounced performance degradation when the tabular‐feature and conditioning pathway was removed, whereas the effects of representation initialization and physics‐informed guidance were target dependent. Comparisons with non‐MEPI baselines showed task‐dependent performance differences, while uncertainty analysis yielded generally near‐nominal LSP interval coverage with modest under‐coverage at higher confidence levels. The frozen model was further implemented in a laboratory prototype for monitoring and frequency screening, supporting the practical feasibility of the proposed framework within the investigated experimental domain.