ABSTRACT
We have developed a system that inspects the entire circumference of tube-to-tube sheet welds using phased array ultrasonic testing for manufacturing of heat exchangers. Therefore, a huge number of images were generated during the inspection. In the past, inspectors made visual judgments based on these images, but due to the long inspection period, it was difficult to inspect all of them. As a solution to this problem, we have developed a system that automatically detects flaws in weld. In constructing the flaw detection system, image processing, AE (Autoencoder), and CNN (Convolutional Neural Network) were examined, but individual judgment accuracy was not sufficient. Therefore, an ensemble method was applied. In the proposed method, the reliability of each discriminator was calculated for each image to be evaluated, and the presence or absence of flaws was finally determined according to the reliability. It was confirmed that the accuracy was improved by applying this flaw detection system.
