Junction Temperature Estimation in IGBT Modules using Machine Learning based Fine-Tuning for Domain Adaptation

Inhalt:
Accurate junction temperature (Tj ) estimation is important for the reliable operation of insulated gate bipolar transistors (IGBTs) in power converters. Traditional thermal models often lack precision when relying solely on manufacturer data, while temperature-sensitive electrical parameters (TSEPs), such as the on-state voltage drop (Vce), require extensive device-specific calibration due to manufacturing tolerances. Machine learning (ML) models trained under fixed conditions typically produce inaccurate results under parameter variations (e.g., Vce shifts). To address this challenge in ML-based Tj estimation, this research proposes a novel supervised fine-tuning (SFT) technique, enabling a model pretrained on one IGBT (Dsource) to adapt to others (Dtarget) with minimal characterization data. Test results on different IGBT power module chips demonstrate that our method significantly reduces Tj estimation errors from 40deg C to approximately 8 °C, achieving a accuracy improvement, thus ensuring reliable Tj estimation under parameter variance.