Sistema imunológico artificial com aprendizagem profunda para detectar defeitos de solda em imagens radiográficas PDVD de tubulações de petróleo

Abstract

In recent years, research focused on (semi)automatic radiographic inspection methods has gained more attention. The present work proposes a method for detecting defects in radiographic images of welded joints of oil pipes. Real condition images obtained by the double wall double image (DWDI) technique usually present a lower quality when compared with images traditionally considered in many studies reported in the literature. First, the proposed approach detects discontinuities in DWDI radiographic images of welded joints, and then, based on a hybrid paradigm encompassing artificial immune systems (AIS) and deep learning (DL), it classifies each discontinuity as ‘defect’ and ‘non-defect’. The proposed method performs two phases in the AIS module: early classification (based on negative selection) and evolving classification (based on clonal selection). In both phases, the pattern recognition task is performed using a set of features extracted from each discontinuity through a detector genetically encoded into immune cells. As an attempt to improve the classification performance, DL models (AlexNet and autoencoders) are incorporated aiming to increase the number of extracted features. Experiments performed on a set of 727 discontinuities show that the proposed approach achieves an F-score of 70.7%, outperforming each of its modules running by themselves: AlexNet with F-score = 64.86% and AIS with F-score = 66%. The experiments also show that the proposed model is capable outperforming tradicional classifiers such as SVM, whose best implementation has achieved an F-score around 60%. Considering the challenges imposed by real conditions on image acquisition and the low rates of false negatives, results demonstrate that the proposed approach can be used to assist experts in their inspection works when dealing with DWDI images.