Weld-Quality Diagnosis of In-Service Natural Gas Pipelines Based on a Fusion Model

Abstract

Damage to pipeline welds during service can result in substantial human and economic losses. Hence, early and accurate quality diagnosis of in-service natural gas pipeline welds is of utmost importance. However, the current diagnostic methods are plagued by low accuracy and excessive resource consumption. Therefore, a fusion model for diagnosing the quality of in-service natural gas pipeline welds is constructed based on gradient boosting decision tree (GBDT), k-nearest neighbor (KNN), and random forest classifier (RFC). Initially, heterogeneous raw data sets undergo data cleaning and quality enhancement processes. Subsequently, the predictions from GBDT, KNN, and RFC are fused using voting as the fusion strategy. The processed data set is then utilized to train and test the fusion model (MIX). After verifying the reliability of the model through cross-validation, the performance of the model’s prediction results is assessed, and a comprehensive analysis of the model’s characteristics is conducted by combining theory and empirical evidence. By analyzing the experimental results, the fusion model performed better than the single model, with an accuracy rate of 96%. It can effectively diagnose the quality of natural gas pipeline welds in service, providing a new reference for existing diagnostic methods. The fusion modeling method proposed in this study can be applied to the diagnosis of welding quality of in-service natural gas pipelines in order to solve the problem of early quality diagnosis of natural gas pipelines and improve the diagnosis of welding quality. The method can effectively overcome the problem of small size and poor quality of engineering data sets. First, the original data set is cleaned and downscaled, and then the synthetic minority oversampling technique (SMOTE) combined with edited nearest neighborhood (ENN) (i.e., SMOTEENN) is balanced to improve the quality of the data set. Then, the data set is predicted by three models, and the three predictions are fused using a voting fusion strategy to effectively improve the diagnostic ability of the models. Finally, the importance of the output features of the fused models is utilized for empirical and theoretical information mining. The experimental results show that the diagnostic accuracy of the method can reach 96% with high reliability and stability, which can provide technical support and theoretical basis for the diagnosis of the welding quality of in-service natural gas pipelines. Mining feature information can also provide more reference information for the pipeline welding construction stage, improve the initial welding quality, and reduce welding defects.