Session: 21-01: Poster Session

Paper Number: 88176

88176 - A Study on Flaw Signal Detection for Phased Array Ultrasonic Testing Using Artificial Intelligence 

The industrial use of flaw interpretation in phased array ultrasonic testing (PAUT) using artificial neural networks in weldments must overcome many challenges. One of the major constraints is the resolution of the PAUT images. In the conventional way of flaw interpretation, inspectors differentiate flaws from non-defect reflectors such as weld boundaries using their vision and experience. In this scenario, differences in display resolution resulting from scan lengths have less impact, since the inspector's experience correct them partially. However, artificial neural networks may perceive PAUT images for the same flaw differently due to differences in resolution. In this study, after constructing the PAUT database from the PAUT data and correcting the resolution, performance of flaw detection using artificial neural networks were checked. The data used in the study were acquired in accordance with the international standard by using a 64-channel phased array ultrasonic transducer of 5 MHz from a defect specimen including cracks, lack of fusions, slag inclusions, porosities, and incomplete penetrations. Using this data, the resolution of the PAUT image is adjusted, and a phased array ultrasonic flaw database is constructed through defect labeling and augmentation. Considering the characteristics of the data, 1-stage object detectors that are not overly complex were used as artificial neural network models.

Presenting Author: Jinhyun Park Sungkyunkwan University