Pipeline leakage is an inevitable problem. However, the small leakage signals of natural gas pipelines are weak, which are inclined to be overwhelmed by strong background noises. The timely and accurate identification of leakage is a crucial and formidable challenge. Therefore, an effective leakage recognition research based on iterative self-updating multivariate variational mode decomposition combined with wavelet energy transform and dual-channel neural network was carried out in this study. Firstly, the MVMD algorithm with self-updating parameters is introduced. Specifically the number of internal modes and the penalty factor are continuously iteratively adjusted through a double-loop strategy, which realizes the adaptive and high-fidelity decomposition of multi-channel infrasonic leakage signals. This effectively avoids the modal aliasing and parameter dependence problems of conventional detection methods. Furthermore, an adaptive continuous wavelet transform enhancement strategy is proposed. Specifically the eigenmode functions are distinguished into high-energy leakage components and low-energy background components by using K-means method, where the enhancement strategy is only implemented for signals of high-energy mode and the feature integrity of low-energy signals is retained.Thereby, the information extraction ability of key features is enhanced. Finally, it is input into the designed dual-channel neural network. Specifically the high-energy channel integrates the attention mechanism and the Max pool to enhance the perception of important features. Meanwhile the low-energy channel utilizes a large receptive field convolution to extract the global background information and perform the fusion and pooling operations for information of different channels, and then utilizes the dual-path collaborative fusion to improve feature perception capabilities. Finally, the experiment was verified. The identification accuracy rate of small leaks (pore size ≤2 mm) reached 97.1%, which was 10% higher than that of the mainstream detection methods. Moreover, it maintained good performance and a short reasoning time in the cross-scenario migration experiment, proving its effectiveness and robustness in the practical engineering applications.