A Pipeline Intrusion Detection Method Based on Temporal Modeling and Hierarchical Classification in Optical Fiber Sensing
Chaoqun Gong, Yiyuan Yang, Haifeng Zhang, Jia Meng, Yunbin Ma, Sibo Du, Yi Li
Abstract
Early detection of intrusion events in long pipelines is crucial for the safe transportation of liquid and gaseous energy sources like petroleum and natural gas. Distributed optical fiber sensing technology has proven to be highly effective for this purpose due to its extensive spatial measurement range and good sensitivity. Also, complex and noisy optical fiber signals can be classified by machine-learning and deep-learning techniques. However, generic network architectures often struggle to capture the unique features of these signals, leading to issues like overfitting and poor generalization. Moreover, with the large volume of optical fiber data, there is a tradeoff between network size and real-time deployment feasibility. In this article, we introduce a novel temporal 2-D modeling approach named OFTNet for distributed optical fiber sensor data. A new representation of optical fiber signals is established, enabling a comprehensive exploration of the spatiotemporal features of the distributed signals. OFTNet excels in high-performance event classification based on optical fiber data. Experimental results show that our approach achieves an accuracy of more than 98% in the data from two deployed energy transportation pipelines and has excellent clustering ability for different events. Furthermore, OFTNet can be compressed to a compact of 10 MB size, meeting real-time recognition requirements and demonstrating great potential for practical application scenarios.