7. Automatic flaw extraction method for nondestructive inspection of bogie parts

Nondestructive inspections such as magnetic particle testing and ultrasonic testing are used to inspect railway vehicle bogie parts. However, determining the presence of flaws from the images or waveforms obtained through these inspections requires experience. Therefore, we have developed a method to automatically extract flaws by machine learning.

In magnetic particle testing of welding surfaces, false indications due to unevenness on the welding surface appear in addition to flaws, so it is necessary to distinguish between the two. As shown in Figure 1, we developed a method that uses a machine learning model to extract areas suspected of having flaws within small, segmented areas. If multiple suspected areas are connected, the method determines them to be a "flaw." The performance of this method was verified using actual magnetic particle testing images, and it was confirmed that surface flaws in weldings could be detected with an accuracy of approximately 70%.

On the other hand, ultrasonic testing is used to inspect the inside of a welding. As shown in Figure 2, the reflected wave from the surface shape of the welding can be confused with the reflected wave from a flaw (flaw echo), so it is necessary to distinguish between the two. In response to this, we developed a model trained on waveforms obtained from simulations of ultrasonic testing of bogie frames using machine learning. We confirmed that this model can detect internal flaws in weldings with an accuracy of over 95%.

By applying the established machine learning method and model to the inspection of actual bogie parts, it is possible to make a uniform judgment on magnetic particle testing images and ultrasonic testing waveforms, from which determining whether a flaw exists or not can be difficult, and to de-skill the flaw detection process.