Life Prediction of IGBT Across Working Condition via a CNN-Transformer Network

Abstract

Insulated Gate Bipolar Transistors (IGBTs) are extensively utilized in a multitude of fields owing to their proficiency in power conversion and their dependable operation. Anticipating the service life of IGBTs to preemptively mitigate the repercussions of device failure, this research advances a novel lifespan forecasting methodology underpinned by a Convolutional Neural Network (CNN) and Transformer hybrid model. The methodology commences with accelerated aging power cycling tests within a range of temperature thresholds, utilizing the Siemens Power Tester to gather aging parameters at disparate junction temperatures. A pivotal observation is the alteration of the saturated voltage drop, VCE(ON), throughout the aging process, which is then harnessed as a critical aging indicator for model training. Following this, the accrued datasets from three distinct groups undergo a rigorous preprocessing phase. Subsequently, the proposed forecasting technique is deployed to predict lifespan across varying operating conditions. The empirical findings underscore that the model introduced in this paper, when predicated on the variations in saturated voltage drop, achieves markedly enhanced predictive fidelity in both single-step and multi-step forecasting scenarios, outperforming alternative comparative methodologies. Especially in single step prediction, the mean values of the coefficient of determination (R2), mean absolute error (MAE), and root mean square error (RMSE) are 0.996, 0.0016 and 0.0026, respectively.