Redes neurais convolucionais profundas aplicadas na inspeção radiográfica de juntas soldadas

Abstract

Periodic inspections of pipelines are carried out mostly based on non-destructive tests and are essential to ensure safety, quality, and reliability in petrochemical facilities. The use of deep learning has gained prominence in different application domains but its use is recent in the context of radiographic inspection of petroleum pipelines. This work proposes a set of methods based on deep convolutional networks to support the activity of radiographic inspection of welded joints in oil pipes. The approaches proposed in this research aim at automatically identifying welding defects that can compromise the structure of installations and equipment under analysis. Therefore, the problem is addressed considering two phases: (i) detection of the welded joint and (ii) semantic segmentation of the welding defects. The detection of the welded joint, whose methodology represents one of the main innovations achieved in the work, identifies and classifies regions on radiographs with Double Wall Double View exposure of welded joints of oil pipelines. Semantic segmentation of defects, on the other hand, locates defects at the pixel level from samples extracted through sliding windows to investigate the compositional inference segmentation method proposed in this work. It is worth highlighting the challenges imposed by the considered dataset, mainly due to the lack of contrast and high level of noise present in most real-world images. The main contributions achieved in this thesis consist of a robustness analysis over different deep convolutional architectures to white and impulsive noise in radiographic inspection and in the proposal of innovative mechanisms for composing inferences to locate the welding defects pixel by pixel. Considering different test groups, the best configurations resulted in an average F-score of 96% when detecting welded joints. When detecting welding defects, the average F-score was 57% in semantic segmentation, i.e., pixel level, and 79% in detection tasks, i.e., object (defect) level.