부산대학교 도서관

Natural gas pipeline leaks may result in huge economic losses, environmental pollution and the loss of life. Therefore researching methods to detect leakage of gas pipeline is an important task in the overall integrity management for a pipeline system. Data-driven machine learning methods have become a popular option, because the methods based on physical model with assumptions are generally unavailable to the real world. However, supervised leak detection fails due to the shortage of leakage fault sample and unbalanced data. To address this problem, this paper proposes an unsupervised leak detection through time-series flow data collected by the Supervisory Control and Data Acquisition (SCADA). The main idea of the proposed method is to learn representative features of flow in the leak-free operation state using the deep auto-encoder (DAE). This method does not require complex feature extraction of leaked data owing to the unsupervised feature extraction scheme. The experimental results show that the proposed pipeline leak detection method using the DAE can provide reliable pipeline leak detection accuracy even without labeling.