GPU-based Multivariate IGBT Lifetime Prediction

In the context of critical energy infrastructures (e.g., hydrogen infrastructure) that extensively utilize power converters, the need for reliable and accurate monitoring is of paramount importance. Addressing this necessity, this paper presents a novel GPU-based multivariate approach to Insulated Gate Bipolar Transistor (IGBT) lifetime prediction. Despite the substantial technological advances in the field, accurately predicting the lifetime of IGBTs remains a significant challenge. Current methods often rely on single precursor variable models, which can lack the precision required in demanding power electronic applications. In contrast, this study utilizes multiple precursor variables (V
CE(ON)
and case temperature) to achieve more accurate results. Initial results using NASA's open-source dataset, and Gaussian Process Regression (GPR) reveal that our multivariate model outperforms its single-variable counterparts in prediction accuracy. Furthermore, the paper elaborates on the development of a small-scale experimental setup to enable real-time health monitoring of the IGBT. It uses an NVIDIA Jetson Nano GPU to handle real-time data from V
CE(ON)
and Case Temperature. A pre-trained random forest model based on V
CE(ON)
and Case Temperature profile data is deployed in NVIDIA Jetson Nano GPU for real-time prediction of the remaining useful life (RUL) of the IGBT, laying the groundwork for future implementation of the real-time multivariate model. The findings from this research highlight the significant potential for enhancing IGBT lifetime prediction accuracy through GPU-based multivariate models.