THMS-Net: A Two-Stage Heterogeneous Signals Mutual Supervision Network for MFL Weak Defect Detection

Abstract

Magnetic flux leakage (MFL) testing is an effective method of pipeline defect detection to guarantee the health of submarine pipelines. However, due to the interference of strong defect signals and complex nondefect signals, the detection of MFL weak defects is a tough issue currently. Focusing on this problem, an MFL weak defect detection method based on two-stage heterogeneous signals mutual supervision network (THMS-Net) is proposed. THMS-Net includes two deep networks for the feature extraction of heterogeneous MFL axial and radial signals, and the mutual supervision of them enhances the discrimination between the defect and nondefect signals. There are two stages in the training phase of THMS-Net. The first stage is to approximate the outputs of THMS-Net with heterogeneous defect inputs, and the second stage is to extend the outputs of THMS-Net with heterogeneous nondefect inputs. THMS-Net establishes a potential relationship between heterogeneous signals through the mutual supervision of defect samples and nondefect samples for the first time. Additionally, an improved nonmaximum suppression (NMS) on external features is proposed to refine defect detection. Finally, the experiments are conducted to verify the effectiveness of the proposed method.