Time Series Anomaly Detection for Natural Gas Pipeline Leakage
Xuguang Li, Zheng Dong, Haobin Zhang
Abstract
Natural gas pipelines play a crucial role in energy transportation, so accurate detection of leak anomalies is vital for safety. Supervisory Control and Data Acquisition (SCADA) systems are widely utilized in the pipeline industry and store extensive historical data with time series characteristics. In this paper, we present a masked Transformer detection model to address the issue of sparse leak labels in SCADA systems and overcome the limitations of neural networks in modeling long-time series. The model incorporates an encoder-only Transformer with a masked mechanism. We validated its effectiveness using real natural gas pipeline data, and the results showed that it can accurately identify pipeline leak anomalies. In particular, compared to other models, the masked Transformer model has shown an improvement in accuracy, recall, precision, and F1 score by 1.4%, 2.5%, 0.3%, and 1.4%, respectively, in real pipeline scenarios. Overall, the masked Transformer model excels in detecting anomalies in natural gas pipeline leakage.