Automated data evaluation in phased-array ultrasonic testing based on A-scan and feature training

In this study, a training dataset and neural network were newly configured to develop a diagnostic artificial intelligence (AI) for phased-array ultrasonic testing (PAUT), and three experiments were conducted to verify its performance. The main problem of this study is to properly classify the suspicious normal signal and defect signal, which is difficult because the welding boundary and geometry signal must be considered together. First, in order to collect high-quality data necessary for training neural network, 40 welding specimens were manufactured and a total of 120 defects were included in the specimens. Then, the neural network was trained using data from 120 defects inspection. Through the first experiment, it was determined that three convolutional layers are sufficient to learn the A-scan signals. In the second experiment, a comparison experiment was performed by adding four features to the training dataset, and it was confirmed that the classification accuracy was significantly improved. Finally, in the third experiment, the trained network was utilized to perform data evaluation for each inspection result file. As a result, the evaluation results by diagnostic AI were mostly consistent with the reference expert's results, while the other expert were unable to properly identify the defect types. In conclusion, the proposed method in this study remarkably improved the classification accuracy of the signals and led to successful data evaluation, which is expected to help the PAUT experts make more accurate data evaluation.