Collapse pressure prediction of mechanically lined pipes using FEM and machine learning techniques

Abstract

Mechanically lined pipe, which proves to be an economical and reliable way to transport corrosive hydrocarbons, has been used in many offshore projects. The presence of liner and the manufacture-induced changes in mechanical properties for both tubes are expected to affect the collapse performance of lined pipes, which, however, are not suitably captured by the current DNV formula. To this end, a dedicated numerical framework, including the axisymmetric and three-dimensional finite element models, was presented to simulate the manufacturing process and the evolution of the external collapse. Two types of commonly used lined pipes, unbonded (LP) and bonded (LPB), were considered. The results indicate that the DNV prediction mostly overpredicts the collapse pressure of typical LP pipes but underestimates that of the LPB cases. Furthermore, the parametric study of essential variables of the problem reveals that in addition to the strong effect of diameter-to-thickness ratio and initial ovality, the impact of the material hardening and plastic expansion history cannot simply be ignored. Thus, the machine learning technique was applied as a substitute resolution to evaluate the critical collapse capacity of lined pipes. Four algorithms, including RF, MLP, KNN, and SVM, were developed based on a dataset of 86,400 cases involving nine variables. Based on the assessment of typical statistical metrics, it was found that all four machine learning models outperformed the DNV formula. Among them, RF had the best prediction performance, while KNN had the lowest precision among the chosen surrogate models.