LAUSR

The defect detection of magnetic flux leakage (MFL) in-line inspection is of great importance for pipeline safety transportation. However, it is difficult for conventional defect detection methods to achieve satisfactory results with limited labeled samples. This article proposes a MFL mechanism-based self-supervised method (MMSM), which can substantially improve the defect detection accuracy and universality with insufficient labeled samples. In contrast to the existing defect detection methods, the MMSM has three main characteristics. First, a hierarchical compression method is proposed, which can enhance the feature expression of weak defect signals to guarantee the accuracy of the detection results. Second, according to the features of MFL data, a self-supervised network based on the MFL mechanism is proposed, in which unlabeled data can be trained to reduce the dependence on labeled data. Third, a novel depth separable deformable convolution (DSDCN) and three-stage online hard example mining (T-SOHEM) head network are proposed, which enhance the accuracy of defect detection and reduce noise interference in MFL data. The real data are utilized to evaluate the performance of the MMSM. The experiments results illustrate that the defect detection model can be effectively established by the proposed method with limited labeled samples.