Synthesizing Labeled Magnetic Flux Leakage Signals for Pipeline Integrity Assessment: A Generation and Evaluation Methodology

Magnetic flux leakage (MFL) detection plays a vital role in pipeline integrity inspection, enabling the identification and assessment of potential defects in oil and gas transportation systems. While machine learning methods have significantly enhanced the speed and accuracy of MFL signal processing, acquiring large, annotated datasets with specific defect categories remains a major challenge due to the high cost of labeling methods. This paper presents a novel diffusion model approach that incorporates class labels to generate synthetic MFL data for specific types of pipeline defects. We develop a comprehensive methodology that includes both simulation modeling and pull through testing to gather training data on various pipeline defects. The quality of the synthetic data is rigorously evaluated through multiple approaches. Visualization results demonstrate that the shallow features of both synthetic and natural datasets are highly similar, while analyses using PCA, t-SNE, and KNN confirm their similarity in deeper feature spaces. The proposed method provides a reliable solution for generating high-quality annotated magnetic flux leakage data, addressing a critical bottleneck in pipeline defect detection systems.