Multi-resolution leak detection based on shared expert MoE forecasting for natural gas pipelines

Download PDF
Open Webpage

Xuguang Li, Zhonglin Zuo, Zheng Dong, Hongke Zhao, Luanfei Wan, Hongfang Cheng

Published: 2026, Last Modified: 06 Oct 2025Inf. Process. Manag. 2026EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Natural gas is a critical strategic energy resource, predominantly transported through extensive pipeline networks monitored by Supervisory Control and Data Acquisition (SCADA) systems. Developing accurate deep-learning models for pipeline leak detection using SCADA data is crucial for safeguarding this vital infrastructure. Reliable and timely leak detection remains challenging due to two inherent limitations: (1) severe sample imbalance from rare leak occurrences and (2) complex multi-resolution hydraulic patterns complicating leak characterization. To address the challenges, we propose a novel Multi-Resolution Shared-Expert Mixture-of-Experts (MR-SEMoE) framework for leakage detection. The framework employs multivariate time series forecasting, where deviations between predicted and observed sensor values trigger leak alarms through statistical thresholding. Two key innovations synergistically enhance detection performance: (1) a shared-expert MoE architecture improving generalization through cross-experts knowledge transfer. (2) A multi-resolution analysis framework featuring parallel multi-head forecasters with resolution-specific feature extractors that enable hierarchical representation learning across different time resolutions. Comprehensive experimental evaluations on real-world natural gas pipeline datasets demonstrate that the proposed MR-SEMoE effectively identifies leaks under imbalanced data conditions. Compared to the previous state-of-the-art method, MR-SEMoE’s F1-score improved by 1.67%. The MR-SEMoE model outperforms contemporary state-of-the-art approaches, establishing the premier natural gas pipeline leak detection framework. To our knowledge, this work constitutes the first successful implementation of the MoE methodology in this domain, facilitating future deployment of large-scale models.