A Novel Machine Learning-Based Strain Capacity Prediction Model of High-Grade Pipeline Girth Welds Using LightGBM

Currently, the non-uniformity of girth weld positions makes their limit state a crucial determinant of pipeline safety. The design method based on the limit state is pivotal in ensuring the integrity and reliability of the pipeline system. Challenges often emerge when determining the limit states of girth welds using semi-empirical formula methods, primarily due to difficulties in accurately identifying influential factors. The quantitative impact of each influence parameter on the crack driving force and the results determined by the semi-empirical formula remain unclear. This study utilizes numerical simulation methods to systematically analyze the quantitative sensitivity laws of critical factors such as crack depth on the crack driving force to address this challenge. The findings revealed that the strength matching coefficient, crack depth, and misalignment are the most significant factors influencing the crack driving force, followed by crack length, softening rate, yield-to-strength ratio, internal pressure, and wall thickness. The effects of tensile strength and outer diameter are relatively minor. A comprehensive database of crack driving forces is constructed using a parameter matrix approach. Combined with the LightGBM machine learning algorithm, a full-scale prediction model for the strain capacity of pipeline girth welds is developed. Predictions for 18 sets of wide-plate test results from the literature confirm the high accuracy of the prediction model, with a prediction accuracy of 6.48%. This research provides a robust reference for accurately determining the limit state of pipeline girth welds and effectively meets the demands of rapidly advancing welding technologies and increasingly complex service environments.