The Effect of Different Flaw Data to Machine Learning Powered Ultrasonic Inspection

Previous research (Li et al., Understanding the disharmony between dropout and batch normalization by variance shift. CoRR abs/1801.05134 (2018). http://arxiv.org/abs/1801.05134arXiv:1801.05134) has shown the plausibility of using a modern deep convolutional neural network to detect flaws from phased-array ultrasonic data. This brings the repeatability and effectiveness of automated systems to complex ultrasonic signal evaluation, previously done exclusively by human inspectors. The major breakthrough was to use virtual flaws to generate ample flaw data for the teaching of the algorithm. This enabled the use of raw ultrasonic scan data for detection and to leverage some of the approaches used in machine learning for image recognition. Unlike traditional image recognition, training data for ultrasonic inspection is scarce. While virtual flaws allow us to broaden the data considerably, original flaws with proper flaw-size distribution are still required. This is of course the same for training human inspectors. The training of human inspectors is usually done with easily manufacturable flaws such as side-drilled holes and EDM notches. While the difference between these easily manufactured artificial flaws and real flaws is obvious, human inspectors still manage to train with them and perform well in real inspection scenarios. In the present work, we use a modern, deep convolutional neural network to detect flaws from phased-array ultrasonic data and compare the results achieved from different training data obtained from various artificial flaws. The model demonstrated good generalization capability toward flaw sizes larger than the original training data, and the effect of the minimum flaw size in the data set affects the a90/95 value. This work also demonstrates how different artificial flaws, solidification cracks, EDM notch and simple simulated flaws generalize differently