Prediction of Corroded Pipeline Failure Pressure Based on Empirical Knowledge and Machine Learning

This paper presents a novel approach for predicting the failure pressure of corroded pipelines by integrating empirical formulas into the loss function of a neural network-based prediction model. Traditional empirical formulas, such as ASME-B31G, DNV RP-F101, and PCORRC, have been widely used for their simplicity but often suffer from significant prediction errors due to the complex interactions between defect parameters and material properties. In contrast, artificial neural networks (ANNs) offer more accurate predictions but require substantial training data. To address these limitations, we propose an integrated loss function that combines the strengths of empirical formulas and the powerful fitting capabilities of ANNs. The proposed loss function incorporates an additional defect factor term predicted by the neural network to compensate for errors caused by varying defect conditions, thereby enhancing the model′s adaptability and accuracy. The model is trained using a diverse dataset of 60 burst test results from various literature sources, covering a wide range of corrosion scenarios. The results demonstrate that the proposed method significantly improves prediction accuracy compared to traditional empirical formulas and ANN models trained with standard loss functions. The proposed approach achieves a mean absolute percentage error (MAPE) of 2.52%, a root mean square error (RMSE) of 0.39 MPa, and a coefficient of determination (R2) of 0.9886 on the validation set. This study highlights the effectiveness of integrating empirical knowledge with data-driven models and provides a robust and accurate solution for predicting the failure pressure of corroded pipelines, contributing to enhanced pipeline integrity assessment and safety management.