Machine Learning Models of Burst Strength for Defect-Free Pipelines

Burst strength of line pipes is essential to pipeline design and integrity management. The simple Barlow equation with the ultimate tensile strength (UTS) was often used to estimate burst strength of line pipes. To consider the plastic flow effect of ductile steels, Zhu and Leis (2006, IJPVP) developed an average shear stress yield criterion and obtained the Zhu-Leis solution of burst strength for defect-free pipelines in term of UTS and strain hardening exponent, n, of materials. The Zhu-Leis solution was validated by more than 100 burst tests for various pipeline steels. The Zhu-Leis solution, when normalized by the Barlow strength, is a function of strain hardening rate, n, only, while the experimentally measured burst strength, when normalized by the Barlow strength, is a strong function of n and a weak function of UTS and pipe diameter to thickness ratio D/t. Due to difficulty of three-parameter regressions, this paper adopts the machine learning technology to develop alternative models of burst strength based on a large database of full-scale burst tests. In comparing to the regression, the machine learning method works well for both single and multiple parameters by introducing an artificial neural network (ANN), activation functions and learning algorithm for the network to learn and make predictions. Three ANN models were developed for predicting the burst strength of defect-free pipelines. Model 1 has one input variable and one hidden layer with three neurons; Model 2 has three input variables and one hidden layer with five neurons; and Model 3 has three input variables and two hidden layers with three neurons for the first hidden layer and two neurons for the second hidden layer. Those ANN models were then validated by the full-scale test data and evaluated through comparison with the Zhu-Leis solution and the linear regression result. On this basis, the best ANN model is recommended.