Development of an integrated deep learning-based remaining strength assessment model for pipelines with random corrosion defects subjected to internal pressures

Accurate and fast estimating the residual strength for corroded pressurized pipelines is crucial for integrity management. Owing to harsh marine environments, realistic corrosion defects for offshore pipelines are random and non-uniform, substantially affecting burst failure behaviours. Addressing this point, based on the random field (RF), finite element analysis (FEA) and convolution neural network (CNN), an integrated residual strength assessment model was developed — through coupling RF and FEA, a theoretical-numerical approach was derived to generate random corrosion morphologies of defects (input) and solve the corresponding residual strengths (output), which subsequently constituted the datasets for training and evaluation of the CNN-based prediction models. The results indicate that, mechanical behaviours during the failure development caused by corrosion morphologies were well captured in the developed models, including stress concentration and redistribution, restrictions to hoop tensile and interacting effects. On this basis, the models showed good performance in predicting residual strengths for both isolated and interacting random defects. Furthermore, detailed influences from related factors on model performance were discussed and explained from mechanics and machine learning principles. Besides, for engineering safety designs, the models exhibited promising capabilities in quantifying the probabilistic characteristics of residual strengths, with an improved computation efficiency of over 30, 000 times. •A CNN-based residual strength pb assessment model of pipelines with random corrosion defects was built.•A RF-FEA model was derived to yield the datasets of corrosion morphologies and pbs.•Failure behaviours caused by corrosion morphologies were well captured in the prediction models.•Model performances for various scenarios were studied, showing good robustness and efficiency.•This study gave a paradigm for target prediction with spatial variabilities using deep learning schemes.