In this study, machine learning models for the estimation of microcrack shape in ferromagnetic materials were constructed using magnetic particle shape data obtained by magnetic particle testing. Although magnetic particle testing detects whether microcracks exist or not, it is difficult to estimate the crack shape, e. g., width and depth, quantitatively. Quantitative evaluation of crack shape enables judgement of material destruction and is of great benefit to the industry. The data of magnetic particle width and height measured in a previous study and the magnetization angle to the crack, which is important in magnetic particle testing, were used. However, since controlling the magnetization angle in magnetic particle testing is difficult in general inspection situations, practical conditions without magnetization angle as input data were examined. Also, measurement data of crack width and depth as target data were used. Machine learning by a neural network was conducted using these data sets, and the effects of estimation accuracy on multilayering of the network were verified. As a result, model performance was improved by optimization of the layer structure on the neural network under each condition, and high performance was obtained for crack depth estimation, which is related to material destruction in particular.
SAKAMOTO et al. (Tue,) studied this question.