Neural network based reliability analysis of pipelines with randomly distributed internal corrosion

Low carbon steel pipelines deteriorate with time and widespread random corrosion of carbon steel has been well reported. The performance of such pipelines is affected by the uncertainties in geometric, material and loading parameters. In understanding the reliability of the pipelines, a clear pathway to failure modelling is essential. The development of closed form analytical functions for determining the characteristic collapse pressures of pipelines subject to randomly distributed internal corrosion defects is difficult because of the geometrical complexities involved. Surrogate models developed with simple regression analysis can sometimes be used but with the disadvantage of comparatively high-cost functions (mean square errors) in some instances. Artificial Neural Networks (ANN) on the other hand can be used to develop more suitable surrogate models which can be readily utilised for reliability analysis.
In this work, response space data of collapse pressure from the finite element analysis of a pipeline with randomly distributed internal corrosion defects is used to develop an ANN model which is applied to perform reliability analysis. The results reveal that the reliability of deteriorating pipelines subject to random corrosion is underestimated if a uniform corrosion is assumed. This method of performing reliability analysis provides a more economical and practical alternative for complex structures than the traditional approaches which require multiple expensive simulations for each analysis instance.