Pipeline defect size quantification via multi-branch online distillation with magnetic flux leakage priors and adaptive fusion

Due to the complex industrial inspection environment, the limited number of labeled defect samples, and the insufficient exploitation of inherent defect characteristics in existing quantification methods, achieving accurate and reliable high-precision three-dimensional size quantification of defects directly from magnetic flux leakage signals remains challenging. To address the above issues, this paper proposes a pipeline magnetic flux leakage (MFL) defect three-dimensional sizing method based on multi-branch online knowledge distillation. This method takes the collected three-axis MFL data as input, extracts information in a targeted manner via multiple branches, and leverages the Magnetic-prior Network (Net-M) branch to effectively fuse prior knowledge and features. The adaptive fusion subnet (AFS) is introduced to adaptively fuse extracted data features, thereby enhancing the model's robustness and interpretability in few-shot scenarios and under practical signal disturbances. In addition, a dual supervised learning mechanism is incorporated to ensure stable model training, and finally, collaborative regression outputs of defect length, width, and depth are obtained. Extensive experiments conducted on the MFL defect data from the pull-through tests of the Φ1016 pipeline show that the average quantification errors of the proposed method are 0.67 mm in length, 1.51 mm in width, and 0.60 mm in depth. The proposed method improves the defect sizing accuracy by >10%, and its detection precision meets the requirements of ultra-high-definition pipeline inline inspection (ILI) equipment. It enables intelligent batch-level quantification of pipeline defect dimensions and maintains high performance with a total dataset size of 50 and across different pipe diameters.