Identification of normal and abnormal from ultrasound images of power devices using VGG16

Abstract

Power devices are semiconductor devices that handle high voltages and large currents, which are used in electric vehicles, televisions, and trains. Therefore, high reliability and safety are required, and to ensure this, power cycle tests are performed to analyze the breakdown process. Conventional tests are often difficult to analyze due to the influence of sparks generated during the test. Therefore, new tests are being developed by adding ultrasound to conventional methods. The new technology is capable of continuously recording structural changes inside the device during testing, which is expected to make testing much easier than conventional testing. However, the new technology still has some challenges. The main problems are the lack of a method for analyzing large amounts of image data and the extraction of small changes in image features that are difficult to distinguish with the human eye, and the establishment of such a system is required. In this paper, we use deep learning for image classification of the obtained ultrasound images. We propose a new network model with the addition of Batch normalization and Global average pooling to VGG16, which is a pre-trained model. In the experiment, accuracy=98.29%, TPR=98.96% and FPR=7.43% classification accuracy was obtained.