A data-driven approach for pipe deformation prediction based on soil properties and weather conditions

The health condition of infrastructure including water transmission and distribution mains has a great impact on the quality of human life. The performance of these water infrastructure is affected by the surrounding soil environment as well as the weather or climate changes. To investigate the structural response of water mains to varying soil movements, field data were collected with a sensor monitoring system. This included pipe wall strain in-situ soil water content, soil pressure, and temperature. Combined with weather factors, an automatic variable selection method, i.e., recursive feature elimination, was first applied to identify critical predictors contributing to pipe deformation. Then, a super learning algorithm was employed to characterize the relationship between pipe deformation and environmental factors. Both base and super learners were built to predict three types of pipe deformation which verified the adaptability of two modeling methods to different predictive models. Predictive performance was evaluated through R-squared, root-mean-square error, and mean absolute error values. The performance metrics demonstrate the advantage of the super learning algorithm in comparison with the baseline methods, especially its capability to further incorporate extra base learners and predictors in a more complex setting. This study shows that the pipe structure behavior could be successfully inferred from surrounding soil properties and weather conditions.