Early Assessment of Pipeline Failure using A Data-Driven CIC Framework

Early assessment of failure events of oil and gas pipelines promotes a faster and more appropriate response plan, which in turn reduces the potential economic and environmental impacts. One of the major challenges in developing a robust failure identification system is the diverse and often incomplete initial failure reports. Different oil and gas pipeline operational entities and service providers may opt to provide incident reports in a shared database. However, they adopt slightly different reporting on a given incident. In addition, the problem is exacerbated when oncoming failure reports are incomplete, which is often the case, and a data-driven model requires a structured feature space to operate. To this extent, this work proposes a scalable semi-supervised Cluster-Impute-Classify (CIC) learning framework which is capable of learning useful clusters of similar incidents as the oncoming report. The matched existing oil and gas pipeline databases are then used to imputer missing information using a tensor decomposition approach. To quantify the magnitude or type of failure, an on-the-fly classifier training is then utilized within a homogeneous ensemble learning environment. The matched and reconstructed descriptors, predetermined from a diverse set, is finally used to classify event types. The proposed CIC framework is shown to effectively reconstruct missing information and identify the failure magnitude when demonstrated on a pipeline failure database from operational entities in the United States.