A Novel Weld Defect Detection Method for Intelligent Magnetic Flux Leakage Detection System via Contextual Relation Network

Accurate weld defect detection is a critical signal processing technique for intelligent magnetic flux leakage (MFL) detection system. However, the MFL signals between the defect and the weld are coupled with each other, and the features between the two are difficult to distinguish, resulting in unsatisfactory accuracy of weld defect detection. To address this issue, a novel weld defect detection method named contextual relation network (CRNet) is proposed to achieve high-accuracy weld defect detection. First, based on the strong coupling relation of MFL signals, a feature extraction network with sparse attention is proposed to extract high-quality features in a sparse manner. Second, a multiscale relation exploration module is proposed to adaptively aggregate multiscale features and fully mine the coupling relation between defect and weld, so that the contextual dependencies between them can be distinguished, and more valid semantic features can be acquired. Third, a multiview collaborative detection mechanism is proposed to realize complementary and coarse-to-fine decisions, so that more accurate detection results can be obtained. The CRNet can clarify the coupling relation and feature differences between defect and weld, guaranteeing independent feature extraction and high detection accuracy. Experiments show that CRNet can reach an average detection accuracy of 95.5%, which is more effective than advanced methods and has strong practicality.