Analysis of Gas Pipeline Failure Factors Based on the Novel Bayesian Network by Machine Learning Optimization

Assessing the failure of urban gas pipelines is crucial for identifying risk factors and preventing gas accidents that result in economic losses and casualties. Most previous studies on gas pipeline failure assessment are based on the basic Bayesian network (BN). However, the conditional probability table (CPT) determined by expert experience is often affected by subjective factors, leading to inaccurate evaluation results. This paper proposes a novel gas pipeline failure risk assessment model based on Bayesian network optimized by machine learning. Firstly, the pipeline fault tree model is constructed according to the accident data that leads to pipeline failure, and the fault tree model is mapped to the Bayesian network. Secondly, by establishing a neural network model, the value of the CPT of the Bayesian network is determined. Aiming at the problem that the true probability of a Bayesian network child node cannot be obtained, the Noisy-OR gate model is introduced to determine the CPT of the node. The advantage of the novel Bayesian network model is that the value of CPT is determined based on data-driven, which effectively reduces the subjective impact of expert experience. Subsequently, the pipeline risk analysis is conducted based on statistical data from the Pipeline and Hazardous Materials Safety Administration (PHMSA). The two-way inference function of Bayesian network is used to obtain the failure probability of gas pipeline and the posterior probability of root node. By comparing with the results of traditional methods, the rationality of the results obtained by the model is verified. Finally, the sensitivity analysis identified Other (Other external damage), third-party excavation damage, incorrect operation and fire or explosion as key influences on gas pipeline failure. The Bayesian network model established in this study is based on data-driven. It effectively reduces the effect of subjectivity brought by expert experience and helps to improve the accuracy of failure assessment. This study reflects that machine learning technology will be an important research direction in the field of gas pipeline risk assessment in the future.