Synthetic Magnetic Flux Leakage Signal Generation Using Diffusion Models: A Novel Approach to Improve Pipeline Defect Detection Accuracy

Magnetic flux leakage (MFL) based inline inspection techniques effectively ensure the safety of oil and gas pipeline transportation. The application of deep learning models to MFL signal processing facilitates rapid and precise defects detection. However, the effectiveness of these data-driven methods is limited by training dataset size, resulting in suboptimal performance. To address this issue, this paper introduces a data augmentation strategy utilizing diffusion models to generate MFL data for training. Specifically, a U-Net architecture is trained to predict noise at each denoising iteration. Synthetic MFL data is produced by progressively removing predicted noise from randomly initialized Gaussian noise. Additionally, category information is encoded into the network to enable the generation of category-specific MFL data. The proposed method is validated using actual MFL data obtained from pull-through testing. Visualization and feature analysis results demonstrate that the synthetic MFL data matches the distribution of actual MFL data. Comparative experiments confirm that the generated MFL data significantly enhances the performance of downstream classification and detection tasks, highlighting its positive impact on accelerating convergence and stabilizing the training process.