A small leakage diagnosis method for natural gas pipelines based on time–frequency self-contrastive learning

Abstract To address the challenge of accurately diagnosing small-scale leaks in natural gas pipelines, a novel diagnostic method based on time–frequency self-contrastive learning is proposed. First, the acoustic emission signal is denoised using the Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise, whose parameters are optimized via the Newton–Raphson Broyden Optimization algorithm. The denoised signal is then fed into a dual-branch time–frequency convolutional neural network to extract features from both the time and frequency domains. Additionally, a cross time–frequency information enhancement strategy is proposed to promote more effective fusion and interaction between domains. At the same time, a new time–frequency information self-contrastive learning strategy is introduced to enhance the model’s ability to learn domain-invariant representations by constructing self-contrastive pairs using homogeneous and heterogeneous transformations across time and frequency domains. Finally, the diagnosis of leakage aperture is achieved through the softmax function. Field experiments have shown that the accuracy of this method in small leak diagnosis is 95.6%. Compared with the existing methods, the method proposed in this paper has higher accuracy than several representative methods for diagnosing small pipeline leaks.