Abstract Inter-turn short circuit (ITSC) faults are among the most critical and frequent failures in power transformer windings. However, conducting a quantitative analysis of the winding insulation state based on MFL remains challenging. This paper proposes a magnetic-electrical spatial state model that links local fault currents to leakage magnetic field variations. A data-driven fault localization framework is developed by combining recursive feature elimination (RFE), Spearman correlation analysis, and support vector machine (SVM) classification. Experimental validation on a 3 kW dry-type transformer, enhanced with FEM-based signal augmentation, shows that the method achieves 97.4% fault localization accuracy under rated load using only 20 Hall-effect sensors. Under no-load conditions, the accuracy remains 92.3%, demonstrating robustness against weak excitation and electromagnetic noise. The optimized sensor layout in the winding gap enhances spatial sensitivity while minimizing hardware complexity. These results confirm the method’s potential for scalable, non-intrusive insulation monitoring in practical power transformers.