In power electronic systems, the variation in on-state resistance of the power switching device metal-oxide-semiconductor field-effect transistor (MOSFET) serves as a key indicator reflecting the device′s health condition, power loss level, system operating efficiency, and reliability. To address the problems of low accuracy, insufficient utilization of topological structure information, and weak model interpretability in existing MOSFET on-state resistance parameter identification methods under complex coupling operating conditions, this article proposes a two-stage hybrid model for multi-parameter identification by fusing Kolmogorov-Arnold networks (KAN) and heterogeneous graph Transformer (HGT). Firstly, time-domain statistical features and frequency-domain wavelet packet local energy values are extracted from the output voltage signal. A combined feature optimization strategy based on the Pearson correlation coefficient and mutual information is utilized to select features with high sensitivity and low redundancy as the key feature subset. Secondly, multiple parallel KAN sub-networks are established to realize the preliminary nonlinear mapping of on-state resistance for each MOSFET. Finally, a heterogeneous graph reflecting the physical semantics of the circuit is constructed, and topologyaware structured compensation is performed on the preliminary identification residuals through HGT to capture the nonlinear characteristics of devices and the coupling relationship of the system, so as to obtain the final identification results. Experimental results show that the average relative error of the proposed method for MOSFET on-state resistance identification in the three-phase space vector pulse width modulation (SVPWM) rectifier system is 0.962%, which is superior to traditional neural networks and homogeneous graph models. The average relative error of identification in hardware experiments is 1.992%, showing favorable generalization and robustness. Finally, this article verifies the effectiveness of feature selection and the physical consistency of model decision-making through Shapley additive explanations (SHAP) analysis and attention visualization, which provides a new idea for the high-precision and interpretable intelligent diagnosis of power electronic devices.