Detection of Multiscale Deterioration from Point-Clouds of Furnace Walls

Deterioration surveys of large structures such as furnaces have been mainly conducted by visual inspection, but it is desirable to automatically detect deterioration using point clouds captured by the terrestrial laser scanner. In this study, we propose flexible methods for detecting various scales of cracks, delamination, and adhesion on furnace walls by using a machine learning technique. Since small cracks have few geometrical features, they are detected from the reflection intensity images generated by projecting a point cloud onto a two-dimensional plane. For detecting cracks on the image, we use the U-Net fine-tuned by crack images denoised with a median filter. For detecting delamination and adhesion, a wall surface is approximated by a smooth B-spline surface, and deterioration is detected as differences between the point cloud and the approximated surface. However, in this method, the resolution of the B-spline surface has to be carefully determined according to the expected deterioration sizes. To robustly detect deterioration at various scales, we introduce multiscale 3D features, and detect deterioration using both multiscale 3D features and 2D features. In actual walls, it is difficult to distinguish between cracks and delamination because delamination grows from cracks. To detect both types of deterioration in a uniform manner, we combine the two detectors and propose an integrated detector for detecting deterioration at various scales. Our experimental results showed that our methods could stably detect various scales of degradation on furnace walls.