Intelligent Optical Microscopy Defects Assessment of Silicon-Carbide Power Modules embedded in Next Generation Electric Cars

In automotive and industrial application domains, the "health monitoring" or predictive reliability of electronic devices plays a key-role in ensuring reliability of the electrical components. Health monitoring regards a collection of methods aimed at monitoring the operational features and performance of an electronic components in order to schedule preventive/predictive maintenance. Similarly, predictive reliability makes it possible to test electronic devices in an intelligent manner, i.e. by identifying predictive patterns of degradation or anomalous behavior in order to effectively characterize the monitored systems. Specifically, the authors of the present work will show the excellent results obtained by applying artificial intelligence techniques to the monitoring by optical microscopy of the power modules in Silicon Carbide technology usually placed in the traction inverter of electric vehicles.. The Silicon-Carbide Power Module is a key component embedded in the traction inverter-system of the modern electric cars, and its functional degradation or unusual behavior may have a significant negative impact on the car’s safety. As with any electronic components, the power module is subjected to electro-mechanical stresses which impact its functionality over time, highlighting progressive states of physical degradation. One technique for detecting defects in Silicon-Carbide Power Modules is optical microscopy. Through ad-hoc deep learning system based on the use of an enhanced YOLOv5 architecture, the captured optical microscopic image of the analyzed Silicon-Carbide Power Module will be automatically processed providing in output a bounding-box segmentation of the most impactful and significant defect patterns for the functioning of the monitored power device. Several tests have been performed on different power modules of the ACEPACK
TM
DRIVE series delivered by STMicroelectronics. The collected performance (identification and classification accuracy close to 93%) confirmed the effectiveness of the proposed solution.