Anomaly Detection from IGBT Images Based on VGGNet

Power devices are semiconductor devices used for power control. They handle high voltages and large currents and are used in electric vehicles, televisions, and trains. These devices are also used in rice cookers, microwave ovens, air conditioners, etc. In homes, they are used in rice cookers, microwave ovens, air conditioners, and so on. Since power devices are electronic components closely connected to our daily lives, they must be highly reliable and safe. To ensure this, power cycle tests are conducted. During these tests, devices are subjected to electrical and thermal stress by repeatedly switching the power on and off. This simulates actual operation and analyzes the breakdown process of joints and other parts on the device chip. However, conventional tests generate sparks during the breakdown process. This severely damages the diode, which is the main part of the power device. This makes it difficult to identify the cause of the breakdown and analyze the process leading up to it. To solve problems in conventional power cycle testing, a new technology using ultrasonic observation is being developed. This technology makes it possible to output images of the test object in real time. Consequently, it is possible to continuously record structural changes inside the device during testing, which enables identification of failure causes and process analysis. There have been problems with conventional testing. However, there are still some issues to be resolved before the new technology can be used in practice. The main issues are the lack of an established method for analyzing large amounts of image data and extracting small changes in image features that are difficult to discern with the naked eye. In this paper, we propose a method for classifying ultrasound images obtained from tests using deep learning. Our method uses the pre-trained VGG16 model and introduces a new network model with an additional skip connection. This allows for the detection of subtle changes in images. Furthermore, we expand the dataset using CycleGAN and Mixup. These techniques reduce the influence of data bias and enable accurate image classification. In our experiments, we applied the proposed method to 201 ultrasound images and achieved a discrimination performance of precision = 0.9767, recall = 0.8936, and F-measure = 0.9333.