Status diagnosis and feature tracing of the natural gas pipeline weld based on improved random forest model

Abstract

The health state of natural gas pipeline welds usually needs to be judged by experienced experts based on the results of excavation works or internal inspections. Due to the low sensitivity of internal detection method, high cost, slow speed and large amount of information of excavation work, maintenance personnel are difficult to accurately and efficiently evaluate the health status of in-service pipeline weld. This paper proposes a model based on the improved random forest for weld joint status diagnosis and feature tracing as a solution to that issue. Firstly, the raw data is cleaned and downscaled by data preprocessing, variance threshold function and correlation heat map. Then an improved random forest model based on the cross-validation is established to predict and analyze the above data. Finally, the model's scores, the features' importance scores, and scores can be acquired. And the parts of the relevant features are analyzed retroactively combined with experts' experience. The experimental results show that the improved random forest model has an average score of 0.941, which is higher than that of the support vector machine model (SVM), k-nearest neighbor classification (KNN), decision tree and random forest models, the decisions generated from the results can help maintain and exclude services of the weld joint.