Actual Maximum Junction Temperature Estimation Process of Multichip SiC MOSFET Power Modules with New Calibration Method and Deep Learning

Abstract

This paper proposes a new estimation process of the actual maximum junction temperature in multi-chip SiC MOSFET power modules. Temperature-sensitive electrical parameters (TSEPs) are common methods to estimate the junction temperature of power devices, but they have limitations in specifying the actual maximum junction temperature in multi-chip power modules with anti-parallel diodes. Common TSEPs includes body-diode forward voltage or on-resistance of MOSFETs, but it is difficult to distinguish the impact of the anti-parallel diodes and the on-resistance-based method has low sensitivity under low-current calibration. Especially, the TSEPs has difficulty to calibrate the actual maximum junction temperature of a power module, which should solely detect the hottest device in the module. Therefore, a new dynamic calibration method is proposed using fiber optics while simultaneously monitoring factors related to the maximum junction temperature. Four critical parameters were measured from two different combinations: 4-paralleled and 10-paralleled SiC MOSFET power modules with anti-parallel diodes. Because of the abundant amount of the measured data, data-driven modeling was conducted with a deep neural network (DNN). The trained models were evaluated, demonstrating distinctly better accuracy in the actual maximum temperature estimation, for both 4-paralleled (MAE of 0.61 ◦C) and 10-paralleled (MAE of 0.72 ◦C) MOSFET modules.

This article proposes a new estimation process of the actual maximum junction temperature in multichip silicon carbide (SiC) MOSFET power modules. Temperature-sensitive electrical parameters (TSEPs) are common methods to estimate the junction temperature of power devices, but they have limitations in specifying the actual maximum junction temperature in multichip power modules with antiparallel diodes. Common TSEPs include body-diode forward voltage or ON-resistance of MOSFETs, but it is difficult to distinguish the impact of the antiparallel diodes and the ON-resistance-based method has low sensitivity under low-current calibration. In particular, the TSEPs have difficulty to calibrate the actual maximum junction temperature of a power module, which should solely detect the hottest device in the module. Therefore, a new dynamic calibration method is proposed using fiber optics while simultaneously monitoring factors related to the maximum junction temperature. Four critical parameters were measured from two different combinations: four- and ten-paralleled SiC MOSFET power modules with antiparallel diodes. Because of the abundant amount of measured data, data-driven modeling was conducted with a deep neural network (DNN). The trained models were evaluated, demonstrating distinctly better accuracy in the actual maximum temperature estimation, for both four-paralleled mean absolute error (MAE of 0.61 C-degrees) and ten-paralleled (MAE of 0.72 C-degrees) MOSFET modules.