Magnetic flux leakage data augmentation for identification of oil and gas pipeline defects: a physics-guided latent diffusion model

Oil and gas pipeline integrity is essential for safe energy transportation, as reliable defect detection helps prevent leaks, thereby reducing the risk of fires and explosions. Magnetic flux leakage (MFL) testing is widely used in pipeline integrity assessment due to its high detection efficiency. However, experimental MFL data are limited, while existing data generation methods lack physical consistency and exhibit low efficiency. These data limitations hinder the training of deep learningbased defect identification models. To address these issues, this study proposes a physics-guided vector-quantized variational autoencoder diffusion (VQ-VAE-Diffusion) model for generating MFL data with clear structural characteristics and strong physical consistency. In this model, a compressed latent representation of MFL data is learned via the VQ-VAE module and discretized through codebook embeddings, thereby reducing data complexity and improving computational efficiency.The diffusion process is guided by magnetic dipole theory, ensuring that the generated data conform to fundamental physical laws. The generated data are evaluated through visual, quantitative, discriminative, and defect identification evaluations. Visual and metric evaluations demonstrate that the generated data are highly consistent with experimental data in spatial symmetry, peak distribution, and edge characteristics. Discriminative evaluation using ResNet18 shows a 55% average classification accuracy between generated and experimental data. Meanwhile, defect identification evaluation using the dataset augmented by the proposed model improves recall by an average of 13% and precision by 1.8%. The proposed approach improves the reliability of defect identification, supporting pipeline integrity management and the prevention of leakage-related safety incidents.