Near-Neutral pH SCCDA Based on Data Analytics and Machine Learning

Abstract External SCCDA plays a pivotal role in managing the integrity and ensuring the safety of onshore gas pipeline transmission systems. Its synergy with specific (ILI) crack detection tools serves as a supplementary approach to address the primary concerns faced by operators: 1) pinpointing the crack locations and 2) accurately determining their dimensions. Presently, gas pipeline transmission enterprises encounter significant challenges concerning the detection threshold and dimensioning performance within EMAT inspections. Additionally, only a handful of companies have developed tools capable of effectively dealing with circumferential crack-like features, and that too within a limited diameter range. The focus of the current project was to address the first question irrespective of crack orientation. In NTS pipelines circumferential and longitudinal cracks are present together in most of the pipes and a differentiation between the susceptibility to circumferential and longitudinal cracks would not be worthy. In several bends, where one would expect to find circumferential cracks, only longitudinal cracks are found instead. Probably because bending stresses contribute to raise the equivalent stresses and the cracks open only in the orthogonal direction to the principal stress in some cases. While SCCDA Standard Practices aim to encompass the theoretical parameters driving the phenomenon, they often yield imprecise results due to the variability in the susceptibility of pipe spools in the field. This is exacerbated by the multitude of parameters that may overshadow the significance of crucial factors and the undue emphasis placed on parameters controlled during laboratory tests. Obtaining reliable information from field inspections, especially under coating conditions, remains challenging. For these reasons, there is a growing demand for tailored susceptibility models for individual pipeline sections rather than a singular generalized model for the entire pipeline. The project focused on soil variables, pipe geometry, coating conditions, and climatic conditions, key factors that can influence the occurrence of corrosion. For soil variables, the physical and chemical properties of the soil were used (mass density, percentages of clay, silt and sand, coarse fragments, soil organic carbon, soil cation exchange capacity, pH). For climatic variables, accumulated precipitation. For the pipe spool characteristics variables, the properties of vertical curvature, horizontal curvature, length, thickness, and coating conditions were used. To collect these variables, satellite data was used for soil variables, data from meteorological stations for climate variables, data from EMAT pig for coating conditions and Inertial pig for the geometry of tubes. To train the model, excavation data was used (which occurs for several reasons, perhaps to explain better) generating a set of training (70%) and test (30%) data. This data provided valuable information, as it was possible to accurately measure whether cracks existed and the characteristics in which they occurred. In the algorithm training process, 4 versions were tested, Extra Trees, Random Forest, Gradient Boosting and Extreme Gradient Booting, and the one that obtained the best results on the test data was chosen. The project revealed two key findings: Firstly, high elastic stresses during the installation phase were the primary drivers for crack initiation in the pipe body, instead of the traditional residual weld stresses along longitudinal seam heat-affected zones. Secondly, although the current technology highlights the importance of coating conditions for the occurrence of the phenomenon, the categorization of coating lack of bonding by EMAT provided an average qualitative value that sometimes may fail to identify poorly coated pipe spool sections accurately in some sections. The best model was the random forest which obtained an accuracy of 0.74, AUC of 0.79, Recall of 0.76, precision of 0.75 and F1 of 0.76. It was also possible to determine the importance of the variables using the shap method (Shapley Additive Explanations) for models with decision tree and the 7 most important were: pipe wall thickness, pipe length, horizontal curvature, vertical curvature, volumetric fraction of coarse fragments, cation exchange capacity of the soil, proportion of sand particles in the fine soil fraction. As the values of these variables exist for the entire gas pipeline, it was possible to determine the location of the pipes with the highest probability of SCC, with a potential cost reduction in the maintenance process and a reduction in the chances of cracks evolving into structural failures.