Chapter 38 STRUCTURAL RELIABILITY ASSESSMENT OF PIPELINE GIRTH WELDS USING GAUSSIAN PROCESS REGRESSION
-
Published:2020
Download citation file:
ABSTRACT
A Structural Reliability Assessment (SRA) of pipeline girth welds by the use of probabilistic machine learning approach coupled with Finite Element Analysis (FEA) have been used to evaluate the probability of fracture for pipelines subjected to large deformations. Firstly, Design of Experiment (DoE) is performed (using two different methods namely One (factor) at a time (OAT) and Latin Hypercube Sampling (LHS)) to generate samples of input variables (wall thickness, outer diameter, initial crack length and height, weld material), which are used to perform FEA. Based on the DOE, a set of non-linear FEA have been performed to derive the crack driving force (CDF) as a function of applied load (i.e. bending strain). The CDF is characterized in terms of the crack tip opening displacement (CTOD) for girth welds with different combinations of input. Different machine learning (ML) surrogate models (SMs) such as Linear regression, Lasso, Ridge, k-Nearest Neighbor, Random Forest, Support Vector Machine, Neural Network and Gaussian Process regression (GPR) have been compared to select the method which best generalize the CTOD response. Based on this comparison the SM GPR_Mv=5/2have been trained and tested on the generated FEA samples, and then used to predict values of CTOD for input combinations that have not explicitly been simulated by FEA. Then, a Monte Carlo Simulations (MCS) have been used to evaluate the probability of failure by comparing CTOD predictions based on the fitted response surface with a representative failure criteria. The advantage of the methodology is its ability to assess uncertainty based on relatively small sample sizes where analytical limit states functions do not exist.