Application of artificial neural network for the structural integrity assessment of dent in pipelines

Abstract

Dent in a pipelines have been of major concern to pipeline operators for years because its severity cannot be easily determined. For many years dent severity was based on dent depth alone. This has led to unnecessary repairs and removal from service incurring considerable loss in revenue. Studies by researchers have indicated that other factors like pipe geometry, pipe material, dent geometry and pressure cycling could influence the severity of the dent in terms of the fatigue life reduction. Dent severity has been studied using dent depth based assessment, strain based assessment and fatigue assessment . The dent depth over the years has been the major determinant of dent severity. Recent studies have shown that the strain in the pipeline could be a better indicator of dent severity using the static approach. The most common fatigue approach is the stress life S-N approach. This involves extracting stress data either through experimental procedure or finite element analysis and using it with an appropriate S-N curve to determine the fatigue life One of the major challenges faced in S-N fatigue approach today is determining the stress concentration factors (SCF) associated with the dents. These SCFs are used with an appropriate SN curve to calculate the fatigue life. This, over the years and currently is calculated empirically or using finite element (FE) analysis. The cost of running experimental program can be very expensive and numerical analysis can be time-consuming. It is not sustainable to keep using finite element analysis to calculate the SCF associated with every dent. An algorithm is needed to be able to predict strain and SCF without running an expensive experimental program or running an extensive finite element study This Research presents an alternative and a sustainable method for calculating the SCF, the maximum strain and the rerounding depth in pipelines with dent. The method involves gathering a large database of SCFs, strains and rerounding depths through a finite element study on a parametric range of industry standard pipes . These parametric datasets focuses on the effects of pipe geometry, dent geometry, material properties and pressure range on the prediction of the strain and stresses which were not systematically considered by other researchers. These parametric datasets are then used to train an artificial neural network (ANN) that predicts the rerounding depth,maximum strain and the SCF. The ANN presents an accurate and sustainable alternative to the current method used for dent assessment. It’s application would reduce the cost and time taken in assessing dent severity. The accuracy of the ANN is dependent on the amount of training data. In order to create the large database of results, a parametric design language (APDL) was created for easy creation and recreation of models. This parametric design language helped in the creation of 256 FE models which was sufficient enough to create the large database of SCF and other data needed to train the ANN Two types of indenters (Dome and Bar) are used to simulate circumferentially and longitudinally aligned dents. Four different dent depths ranging from 2% d/D to 10% d/D are also simulated to investigate the effect of dent geometry. Four different pipe grades (X46, X65, X80, and X100) are analysed to investigate the effect of pipe materials. Similarly, eight pipes with a different diameter to thickness ratio (D/t) ranging from 18-96 are analysed to investigate the effect of pipe geometry. The pipe is pressured up to 50% and 72% SMYS to investigate the effect of pressure range. The results from this study show that all the investigated parameters influence the results in various ways. Results show that longitudinally aligned dents have higher stress concentrations factors compared to circumferential dents of similar dent depth. Similarly, pipes with higher diameter to thickness ratios D/t have higher stress concentration factor compared to pipes with lower D/t .The FE result was validated with experimental and analytical results and a good correlation was seen with minimal percentage error. The FE results from the parametric study was fed into an ANN model to train the network. The network was trained with different numbers of the processing element and activation function to find the model with the best performance. The ANN prediction gave a good correlation with the FE results