A technology for early warning of leaks in oil and gas storage and transmission pipelines based on deep neural networks and multimodal fusion
Kuirong Liu, Binbin XuAbstract
Leakage from long-distance oil and gas pipelines is difficult to recognize under variable operating conditions, background noise, and limited labels. This study combines learnable variational mode decomposition with parallel one-dimensional convolutions of kernel sizes 3, 5, and 7 to fuse short-, intermediate-, and long-scale features from synchronized pressure, flow, acoustic-emission, vibration, and temperature signals. A ProtoNet-based variational retrospective mechanism fuses current and historical prototypes. A 2500-sample dataset was divided in a 7:2:1 ratio and covered pressures of 0.2–0.5 MPa, flow velocities of 0.5–1.5 m/s, temperature variation, and noise. The method achieved 98.50% accuracy and 0.998 macro-AUC (Area Under Curve).