Convolutional Neural Network for Classification of Material Defects in Pulsed Thermal Tomography Images

We have developed a deep learning convolutional neural network (CNN) to classify material defects in pulsed thermal tomography (PTT) images. PTT is a non-destructive imaging technique that uses heat diffusion to visualize subsurface internal defects in materials. In particular, PTT has been investigated for applications in imaging of subsurface pores in metals produced with laser powder bed fusion (LPBF) additive manufacturing method. PTT method involves delivering a thermal pulse on material surface with a flash lamp, and recording surface temperature transients with a fast frame infrared camera as heat diffuses into the material bulk. Thermal tomography reconstruction algorithm visualizes internal structures by converting time-dependent measurements of surface temperature into thermal effusivity spatial depth profile. However, interpretation of PTT images is not trivial because of blurring of images with increasing depth. We address this by developing a CNN for classification of size and orientation of subsurface defects in PTT images. Performance of CNN was investigated using PTT images created with computer simulations of heat transfer in metallic structures. We trained CNN on a database of simulated PTT images of structures with elliptical subsurface air voids, and demonstrated the ability of CNN to classify radii and angular orientation of voids in test images. In addition, we showed that CNN trained on elliptical defects is capable of classifying irregular-shaped defects. Simulated PIT images of such defects were obtained using shapes of actual defects in scanning electron microscopy (SEM) images of sections of stainless steel specimens printed with LPBF method.