Identification of cracks in pipelines based on machine learning and deep learning

Abstract

Pipelines are important long-distance transportation structures in modern industry, and because many are buried deep underground, pipeline health monitoring is critical to industry; however, inspecting underground pipelines can be quite challenging due to the large financial and human resources required. For decades, different methods have been used to assess pipeline cracks. Ultrasonic quantitative nondestructive testing (QNDT) is one of the frequently used methods in pipeline health monitoring. In the current study, the coefficients of the reflected and transmitted waves due to different incident waves were first generated by using a semi-analytical finite element method based on classical elasticity theory. In that study, different types of pipes, including different geometries and materials, were considered. Then four different regression machine learning algorithms and three deep learning algorithms were used to identify crack features. In this study, the prediction accuracy was compared between the different algorithms and different datasets. The objective was to find the algorithm with the highest prediction rate and to select a suitable dataset for prediction. It was found that the extremely randomized tree (ERT) algorithm was the best in identifying cracks in the pipeline. The prediction accuracy will be improved by selecting different data sets. In addition, all algorithms performed better in predicting the radial crack depth (CDRD) than predicting the circumferential crack width (CWCD).