Assessment of arresting performance of integral buckle arrestors for sandwich pipes using machine learning techniques

Integral buckle arrestors are regarded as the most effective arresting devices and can be perfectly adapted to innovative sandwich pipes. In the present study, hyperbaric chamber tests were performed on reduced-scale sandwich pipe specimens equipped with integral arrestors, and the effect of interface bonding behaviour on the crossover pressure was examined. Then, numerical frameworks were proposed to replicate buckling propagating and crossing phenomena under hydrostatic pressure, with a strong consistency between measurements and predictions. A broad parametric analysis on the crossover pressure was implemented covering key material properties and geometries. After that, machine learning techniques were introduced and used for predictions of crossover pressure and arresting efficiency. Four algorithms, involving Random Forest, Multi-layer Perceptron, K-Nearest Neighbors, and Support Vector Machine, were established using a dataset comprising 248 cases with thirteen variables. Based upon an evaluation of standard statistical metrics, it is observed that RF and MLP exhibit superior prediction accuracy, whereas the prediction performance of KNN is the worst. The results show that the machine learning method provides relatively reliable predictions of crossover pressure and arresting efficiency for both flattening and flipping failure modes.