GSA-CYCLEGAN: a global shuffle and aggregation generative network based on mutual fusion and reconstruction of MFL defect signals

Abstract The internal detection of magnetic flux leakage (MFL) serves as a principal non destructive evaluation method for long-distance oil and gas pipeline integrity monitoring. However, the fluctuating state of the detection probe is prone to causing signal characteristic loss and noise interference, resulting in a 30%–40% reduction in defect identification efficiency and a size quantification error of ±15%. Signal interference also triggers redundant verification processes, causing inspection delays, increased costs, and the risk of missed detections. Therefore, this study proposes an improved MFL signal reconstruction method—GSA-CycleGAN. This method is based on the traditional CycleGAN and enhances the feature extraction ability through channel reorganization, shuffling, and aggregation techniques. It also introduces the global second-order pooling attention mechanism to optimize the covariance and improve the detail restoration degree. Experiments were verified using a large-caliber pipeline defect dataset, covering samples with missing three-axis signals and noise interference. The results indicate that compared with the original CycleGAN, GSA-CycleGAN exhibits excellent performance in MFL signal reconstruction: the learned perceptual image patch similarity(LPIPS) is improved by up to 17.14%, the Fréchet inception distance (FID) is improved by up to 19.4%, and the structural similarity index reaches up to 0.97; when processing noise-interfered signals, the peak signal-to-noise ratio is improved by 5.7%, the FID is improved by 14.6%, and the LPIPS is improved by 24.7%. This demonstrates that GSA-CycleGAN can effectively reconstruct signals with data loss and noise interference, significantly enhancing the detection efficiency and reliability.