Long-Distance Pipeline Safety Early Warning: A Distributed Optical Fiber Sensing Semi-Supervised Learning Method

Abstract

Pipeline safety early warning (PSEW) systems based on distributed optical fiber sensors are used to recognize and locate third-party events that may damage long-distance energy transportation pipelines and are essential to ensure pipeline safety and energy supply. However, the deployment of PSEW systems in real sites is hindered by the high experimental cost of collecting large real-site data sets for model building and the small percentage of labeled data (typically less than 0.5%). Besides, the optical fiber sensors are sensitive to hardware and the environment, ensuring challenges to directly migrate the old PSEW system for a new deployment. In this study, a novel semi-supervised learning model is proposed to monitor the safety of pipelines in real-time. Concretely, the sparse stacked autoencoder trained with unlabeled data is used to extract more robust features, and the fully-connected network trained with a small amount of labeled data is used for location and identification. Encouraging empirical results on the real-world long-distance energy pipelines of the PipeChina confirm that our method achieves better recognition and localization performance in comparison to the baseline with less labeled data. Further, the model size and recognition latency are reduced by 18.9× and 7.9× of the baseline, respectively. Also, the decoded features have better visualization than the input. This work reduces the cost of PSEW system deployments, improves its performance and portability, and will contribute to the widespread use of PSEW systems in the industry.