Abstract:Abstract Oil and gas pipeline failure and leakage can seriously damage people's lives and the ecosystem. The prediction of failure pressure for pipelines with damage is one of the most important and challenging tasks faced by industry, which affects the assessment of pipeline safety. Previous studies widely used industrial models or the finite element (FE) method to predict the failure pressure. However, the industrial models may give limited information, and the FE method has much heavy computation burden. In this work, three machine learning models - artificial neural network (ANN), XGBoost (XGB) and CatBoost (CAT) are developed for forecasting the failure pressure of pipelines with defects. Firstly, the simulation results of the FE method are validated by real failure pressure and compared with the calculation results of industrial models (ASME-B31G and DNV). Then 180 pipeline samples including pipeline attributes and defect sizes collected from real in-line inspection data in a pipeline company and the corresponding FE simulation results of failure pressure of these 180 defective pipelines are used for the training and testing of the machine learning models. The results show that the simulation accuracy of the FE method is higher than the calculation accuracy of the industrial models, and the FE simulation results are suitable to be the outputs of machine learning models. The three machine learning methods all provide satisfactory prediction accuracy in failure pressure. Specifically, CAT is the best machine learning method in this study for its lowest relative error (3.11% on average), mean absolute error (0.53), root mean square error (0.78) and highest coefficient of determination (R2) up to 98% in testing. Moreover, the machine learning models present better performance on average relative errors compared to the industrial models. CAT shows higher accuracy than the industrial models and FE simulation on minimum and average relative errors. Finally, the prediction result of CAT is used to discuss the effect of input features on failure pressure of pipelines, which demonstrates that the importance of features follows the order of pipeline thickness > pipeline outside diameter > defect depth > defect length > defect width. Once the above machine learning methods are used in industry, more and more real data will be collected to train a model and make it more accurate. In this way, these methods will provide an efficient way to evaluate the safety of defective pipelines. In addition, the failure pressure of pipeline could be estimated to help operators figure out a pipeline condition and further prioritize the pipelines for maintenance.

Residual Strength Prediction of Pipeline with Single Defect Based on SVM Algorithm

Application of Improved Support Vector Machine in Predicting Failure Pressure of Oil and Gas Pipelines with Internal Corrosion Defects

Probabilistic model to predict longitudinal fracture failure of buried pvc pipes

Novel Data-Driven Framework for Predicting Residual Strength of Corroded Pipelines

Failure Pressure Prediction of Defective Pipeline Using Finite Element Method and Machine Learning Models

PIPELINE DEFECT PREDICTION USING SUPPORT VECTOR MACHINES

Forecasting pipeline safety and remaining life with machine learning methods and SHAP interaction values

Failure risk analysis of pipelines using data-driven machine learning algorithms

Risk Assessment of Submarine Pipelines Based on Multi-Source Data Model

Interpretable Machine Learning Models for Failure Cause Prediction in Imbalanced Oil Pipeline Data

Prediction Model for Pipeline Pitting Corrosion Based on Multiple Feature Selection and Residual Correction

Failure Risk Prediction Model for Girth Welds in High-Strength Steel Pipeline Based on Historical Data and Artificial Neural Network

Model for the Failure Prediction Mechanism of In-Service Pipelines Based on IoT Technology

Mechanical model and probability analysis of buried pipelines failure under uniform corrosion

Rapid failure risk analysis of corroded gas pipelines using machine learning

Data-Based Line Trip Fault Prediction in Power Systems Using LSTM Networks and SVM.

Numerical Simulation and Analytical Prediction of Residual Strength for Elbow Pipes with Erosion Defects

A novel stacking ensemble learner for predicting residual strength of corroded pipelines

Water Pipe Failure Prediction Using AutoML

XGBoost Algorithm-Based Prediction of Safety Assessment for Pipelines

Evaluation of Corrosion Residual Life Prediction Methods for Metal Pipelines