Machine Learning Based Sensor Fusion for Junction Temperature Estimation

Accurate junction temperature estimation is crucial for optimizing SiC power module performance and reliability. A viable alternative to traditional thermistor-based approaches is to utilize temperature-sensitive electrical parameters to implicitly determine junction temperature. In this work, a novel temperature acquisition concept based on the combination of multiple parameters is proposed, using the intrinsic parasitic inductance between the power source and Kelvin source terminals to acquire virtual threshold voltage and virtual time delay during the turn-on and turn-off transitions. Experimental results reveal an improvement of the temperature inaccuracy from ±10K to ±8K using a random forest regression model, offering a promising real-time junction temperature estimation solution for SiC power modules.