Estimating of IGBT Bond Wire Lift-Off Trend Using Convolutional Neural Network (CNN)

Abstract

The health monitoring prediction of power devices is vital for power electronics applications such as renewable converters, electric vehicles, and machine drives. One significant failure mode in the power cycle degradation of Insulated Gate Bipolar Transistor (IGBT) modules is bond wire lift-off. This study uses the gate voltage waveform ( $V_{ge}$ ) as an input to an artificial intelligence (AI) model with the Convolutional Neural Network (CNN). The CNN was demonstrated to accurately estimate the IGBT bond wire lift-off, categorizing it into four levels: no damage, light damage, medium damage, and heavy damage. The Digital Gate Driver (DGD) IC was implemented to generate the $V_{ge}$ and collect the data waveforms by two switching modes: Conventional Vector Control (CVC) and 2-step Vector Control (2-sVC). The experiment evaluated the accuracy of the four-level estimation in several aspects. These aspects include switching modes, the number of datasets, and parts of the waveform The results show that the CNN model achieved high accuracy in estimating the wire lift-off trend. The $V_{ge}$ waveform generated by the 2-sVC switching mode showed better estimation accuracy compared to the CVC mode. Furthermore, it also obtained an effective switching performance $E_{loss}$ - $V_{ce-surge}$ Trade-off curve. Therefore, the DGD is suitable for application and useful for health monitoring and achieving effective switching performance.

Keywords

• IGBT power module
• health monitoring
• power cycle degradation
• bond wire