Leak Localization and Orifice Diameter Estimation in Coal Mine Gas Drainage Pipelines Using a Physics-Informed Dual-Stream Decoupling Network

Leak localization and orifice diameter estimation involve different spatiotemporal scales, complicating joint prediction. This study aims to improve both tasks by constructing a multi-condition dataset from pressure, flow rate, and oxygen concentration signals and proposing a physics-informed dual-stream decoupling network (PI-DSDN). Its transient localization branch extracts multiscale pressure disturbances and uses a pressure-gradient energy mask to focus temporal attention on leakage transients. Its steady-state sizing branch decouples fast and slow variables and uses a macroscopic flow-rate difference to characterize post-leakage redistribution. On the test set, PI-DSDN achieves R² values of 0.9754 for localization and 0.9566 for sizing, with MAEs of 0.283 m and 0.0736 mm and RMSEs of 0.445 m and 0.1129 mm, respectively. Compared with CNN-LSTM, PI-DSDN reduces the localization and sizing MAEs by 45.2% and 45.1%, respectively. Ablation and attention analyses confirm that the physical priors selectively enhance transient localization and steady-state sizing, supporting more interpretable leak assessment.