Machine Learning of Ultrasonic Phased-Array Images for Flaw Detection

Abstract

Ultrasonic phased array is a phase composite imaging technology developed in the radar field and has recently been used for nondestructive inspection of power generation equipment. However, scattered waves in the inspection target make it difficult to distinguish a flaw from scattering from the edge surface. In this study, we developed a method to discriminate flaws with high accuracy by adjusting parameters such as output voltage and receiver sensitivity to make it easier to see the flaws and by deep learning of the optimized flaw images. First, actual measurements were made using an ultrasonic phased-array flaw detector on a stainless-steel specimen. Next, a model was created to discriminate the presence or absence of flaws using transition learning, one of the machine learning methods. As a result, we found that the highest accuracy was achieved when transition learning was performed using inceptionv3 and resnet101, a convolutional neural network architecture. These results show that the method developed in this study is effective for nondestructive inspection.