Research On Neural Network Based Failure Prediction Method For Pipeline Girth Welds

Presently,carbon neutrality and carbon peaking are critical objectives for national development.Enterprises,such as urban thermal power plants,are progressively transitioning their focus from coal to oil and natural gas for energy use.Accurate prediction of the risk associated with oil and gas pipelines is necessary to observe and maintain them as much as possible.Based on the predicted results,high-risk pipelines should receive maintenance.Furthermore,analyzing practical factors has a substantial impact on the prediction of oil and gas pipelines,and it is crucial for construction and pipeline maintenance.This article reviews and analyzes the current research status worldwide to address the above issues.The paper summarizes the main parameters involved in pipeline failure prediction by analyzing commonly used failure pressure formulas internationally.Pipeline risk levels are categorized into high,medium,and low as per the national standard GB 32167-2015 "Oil and gas pipeline integrity management specification".The article mainly focuses on the following aspects:(1)The paper theoretically analyzes and determines a neural network structure suitable for predicting pipeline circumferential weld failure and develops a corresponding pipeline weld failure prediction neural network model.Practical tests demonstrate that the failure risk prediction of circumferential welds after steel pipe classification has the best prediction accuracy of 96.8% when the number of hidden layer nodes is 12.(2)Traditional sensitivity analysis methods are inadequate in comprehensively considering the mutual influence of various production factors in circumferential welds.Therefore,the paper proposes a sensitivity analysis method based on neural networks suitable for predicting the failure of circumferential welds and embeds it into the developed neural network.The calculation indicates that among the factors affecting the risk level of pipeline failure,the importance ranking of each factor is: defect length/pipe diameter > pipe diameter/defect length >> wall thickness > defect depth > yield strength > tensile strength.This conclusion aligns with the observation and summary of actual production.(3)In actual production,there is an imbalance in the amount of data for high,medium,and low risk circumferential welds.To address this issue,this paper proposes a training sample set selection method based on orthogonal optimization.The proposed method constructs a correlation matrix and suggests a training sample selection algorithm that minimizes the overall correlation of double nesting,effectively solving the problem of imbalanced number of various risk samples when training neural networks.(4)Based on actual prediction requirements,the paper designs the functional modules and related database structures of the risk prediction system and develops a failure pressure prediction and sensitivity analysis system based on neural networks using C++.The developed system can automatically predict the risk level of pipeline circumferential welds,effectively assisting operation and maintenance personnel in implementing targeted maintenance measures.