Prediction of Pipe Girth Weld Residual Stress Profiles Using Artificial Neural Networks

Defect assessment procedures such as BS7910, R6 and API 579-1 provide bounding profiles that can be used to characterise the residual stresses present in a weld. The bounding profiles in BS7910 and the R6 procedure have been based on examination of residual stress measurements and expert judgment. This approach suffers from the drawback that the upper bound curve can increase as more measurements and data scatter are obtained. The consequence of this is that structural integrity assessments of defective plant can be over-conservative by a large margin, and may lead to unnecessary and costly repair or inspection. This paper presents work exploring the use of artificial neural networks to predict more realistic residual stress profiles in austenitic stainless steel pipe girth welds. The network is trained using a set of baseline experimental stress data and then tested using previously unseen data. The committee of networks has been implemented and optimised using a maximum likelihood based algorithm. The neural network is validated by comparing predictions with neutron diffraction residual stress measurements of two girth welds 25 mm thick produced using different welding parameters. The capability of the neural network approach is discussed by comparing predictions with stress profiles recommended in the R6 procedure and its potential for development evaluated.