Image-Driven Deep Learning for Prediction and Analysis of Pipeline Leakage and Burst Events

Addressing the challenge of leakage and burst identification in pressurized pipeline systems, this study investigates the pressure-response mechanisms and localization methods under burst events. A systematic analysis of pressure-wave propagation is conducted to reveal the distinct pressure characteristics associated with different burst modes: instantaneous bursts exhibit pronounced fluctuations with clear reflection structures, whereas progressive bursts show noticeable transitional disturbances and buffering stages. Because pressure-cloud maps simultaneously embed both temporal and spatial information, they provide a more comprehensive representation of the propagation trajectory and energy-attenuation pattern of burst-induced disturbances. Accordingly, a long short-term memory (LSTM) model based on time-series pressure data and another LSTM model driven by pressure-cloud-map features are constructed. Four data-partitioning schemes are applied to systematically compare their prediction performance across multiple burst-location scenarios. The results demonstrate that the cloud-map-driven model consistently outperforms the time-series model in all evaluation schemes, achieving lower errors in the root mean square error and mean absolute error, as well as a higher coefficient of determination ( R 2 ). Under conditions with strong pressure-field interference, the maximum R 2 of the cloud-map model reaches 0.98. Experiments on both instantaneous and progressive burst events further show that the cloud-map model maintains an R 2 above 0.95, indicating superior and more stable localization capability. By contrast, the time-series model is more susceptible to complex reflected waves, leading to larger prediction deviations. In summary, pressure-cloud maps represent an essential and highly effective feature modality for burst localization. Their integrated two-dimensional spatiotemporal structure significantly enhances the ability of deep-learning models to accurately identify burst positions.