SERVICE LIFE PREDICTION FOR PIPELINES USING MACHINE LEARNING TECHNIQUES

Abstract

Pipelines are a vital part of the global economy as they transport large volumes of fluids across thousands of miles. Pipelines mainly pass through rural areas and are most of the time buried underground or laid down on the seabed. They are prone to various types of failure due to the harsh environmental conditions they serve in, where corrosion is considered one of the major causes. This study focused on predicting the remaining service life of pipes, before they fails due to corrosion, using a big dataset published by the Pipeline Hazardous Material Safety Administration (PHMSA). The PHMSA dataset includes a large number of predictive fields, along with additional weather data such as temperature and precipitation that were extracted from the National Centers for Environmental Information. A regression model was built to predict the remaining time till failure and classification models to predict an interval of the remaining time till failure. Top performance was achieved using the Extremely Randomized Trees (Extra Trees) algorithm with an R-Squared score of 90.35% for regression and an f1 score of 85% for classification. The importance of each feature used to build the model was assessed using SHapley Additive exPlanations (SHAP) to explain the outputs of the models and identify the most contributing factors responsible for accelerated pipe failure. It was concluded that weather conditions like temperature and precipitation play a major role in pipe failure. Future work may include implementing predictive maintenance using the precise predictions of the time left before failure, as well as considering other datasets from various types of structures.