(185b) A Comprehensive Causation Prediction Model of Pipeline Incidents Using Artificial Neural Network

Failure of oil pipelines results in pipeline incidents that potentially pose significant hazards to people, property, and the environment [1]. The causes of oil pipeline incidents can be broadly classified into five categories: corrosion, equipment failure, incorrect operation, natural forces, and excavation, and each of these causes can be further divided into several sub-causes [2]. To predict the cause and sub-cause of HL pipeline incidents, several data-driven models based on neural networks, regression techniques, and Bayesian method have been developed [3, 4]. However, these approaches utilize only a few data fields from a wide pipeline incident database for the causation prediction of pipeline incidents. Additionally, the existing models are capable of causation prediction for pipeline incidents resulting from only a particular cause. Using cause-specific models, the causation prediction of an impending pipeline incident becomes difficult since the cause is not known in advance. Therefore, there is a need for a comprehensive model which can be utilized for the causation prediction of generic pipeline incidents (i.e., resulting from any cause).

In this work, a comprehensive model for causation prediction of pipeline incidents is presented which leverages the rich pipeline incident database. Specifically, from the pipeline incident data collected by the Pipeline Hazardous Material Safety Administration (PHMSA) in the USA for 2010-2019, 108 data fields have been selected due to their significance for pipeline incidents. These data fields are selected based on domain knowledge to increase the generalizability of the proposed approach. Among the selected data fields, there are two types of data fields: generic data fields that are significant to pipeline incidents resulting from all causes, and data fields specific to a cause that are significant to pipeline incidents resulting from a particular cause. Utilizing the generic data fields, an artificial neural network (ANN) model has been developed to predict the cause of a pipeline incident. Next, for each category of cause, an ANN model has been developed to predict the sub-cause of the pipeline incident utilizing the generic data fields and data fields specific to the cause. Here, although the sub-cause prediction models are specific to a particular cause, the causation prediction model including the cause and sub-cause prediction models is not specific to a particular cause and is applicable to all causes. The effectiveness of the comprehensive model including the cause and sub-cause prediction models in predicting the causation of pipeline incidents is demonstrated using the PHMSA incident database.

Keywords: pipeline incidents; causation prediction; artificial neural network; PHMSA incident database

References:

1. Guo, X. Meng, D. Wang, T. Meng, S. Liu, and R. He, Comprehensive risk evaluation of long-distance oil and gas transportation pipelines using a fuzzy Petri net model, Journal of Natural Gas Science and Engineering, 33, 18 - 29, 2016.
2. S. El-Abbasy, A. Senouci, T. Zayed, F. Mirahadi, and L. Parvizsedghy, Artificial neural network models for predicting condition of offshore oil and gas pipelines, Automation in Construction, 45, 50 - 65, 2014.
3. Senouci, M. Elabbasy, E. Elwakil, B. Abdrabou, and T. Zayed, A model for predicting failure of oil pipelines, Structure and Infrastructure Engineering, 10, 3, 375 - 387, 2014.
4. Breton, J. Sanchez-Gheno, J. Alamilla, and J. Alvarez-Ramirez, Identification of failure type in corroded pipelines: A Bayesian probabilistic approach, Journal of Hazardous Materials, 179, 1, 628 -634, 2010.