Prediction of rupture and perforation limits of pressurised X80 pipelines using BP neural networks and generalised additive models

Abstract

Impact resistance of in-service pressurised X80 pipelines is one of the top safety priorities in the petrochemical industry. Based on impact limits, including the rupture and perforation limits, the failure threshold can be identified and the impact resistance of target structures can be assessed. In this study, backpropagation neural networks (BPNNs) and generalised additive models (GAMs) were developed to predict the rupture and perforation limits of pressurised X80 pipelines. First, an explicit dynamic model was established to provide comprehensive training datasets considering the effects of various dominant parameters. Interdependencies between the impact limits and various dominant parameters were analysed using the Pearson correlation coefficient. Then, BPNN models optimised by the backpropagation algorithm and neuron numbers were trained to predict the impact limits. Based on the optimal smooth functions and smoothing coefficients, the GAMs then provided two integrated regression functions for impact limits. Finally, the predictive stabilities of the BPNN and GAM on impact limits were discussed using a random dataset. The BPNNs and GAMs were in a strong agreement with the expectations, and the predictive accuracy of the BPNNs was approximately three times that of GAMs. The findings of this study can be used as guidelines for protection design and risk assessment of pressurised X80 pipelines.