Technology for detecting small-aperture leaks in natural gas pipelines utilizing transfer learning methodologies

Abstract To address the challenge of limited training samples in natural gas pipeline leakage detection, a novel transfer learning framework is proposed, which requires only a small amount of leakage data from a specific leak aperture as the source domain training set. Additionally, a new residual structure, named Dual-Pieces Net, is designed. This structure combines the cross-layer fusion of residual networks with a convolutional fragmenting mechanism. By processing feature maps in various ways through fragmentation, it enhances the model’s ability to refine and diversify signal processing. Unlike traditional detection models that rely on large amounts of diverse leakage data, the proposed transfer learning (TL) framework deepens the model’s extraction of invariant leakage features from the signal through stacked Dual-Pieces Net modules. This improves the TL performance and provides a solid technical foundation for more diverse TL scenarios. All transfer experiments were conducted on an experimental transport pipeline with a length of 169 m. The experimental results demonstrate that the proposed method offers significant advantages compared to other methods, proving its effectiveness and practical relevance in real-world applications.