基于机器学习的LNG泄漏事故致因分析

Analysis of the accident causality of the LNG leakage based on the machine learning mechanism

为避免泄漏事故发生,以多起LNG泄漏事故调查报告为基础,寻找系统泄漏原因,归纳出以物质状态、设备设施、作业现场、安全管理和应急救援为主的LNG泄漏事故致因体系。将多家LNG储运企业作为正样本,并以多家发生过事故的LNG储运企业为负样本,组成LNG泄漏数据集。对数据集首先运用主成分分析法进行前两主成分分析,绘制出样本散点图,考察前两主成分对分类效果的影响。其次通过对前五主成分进行分析,发现在每一主成分下各特征的权重排序,即对事故发生的影响程度。然后通过随机森林方法进行分类研究,计算出各指标对分类效果的权重。最终通过主成分分析法前五个特征与随机森林权重值的相互比对,分析得出LNG储运企业泄漏事故的主要原因。结果表明,其主要原因集中于以安全管理和应急救援为主的二阶致因指标。

This paper is to devote itself to proposing a machine-learning model based on the principal component analysis(short for PCA)and the random forest classification approach to analyze the sort of importance in the constituent factors likely to cause the LNG leakage accidents.Based on the pattern recognition and statistical learning method,we can find the main factors leading to the LNG leakage by analyzing the LNG storage and transportation data,which is of great significance to improve the overall safety of LNG storage and transportation.And,the first step in the given research is to establish a system of the accident causality of LNG leakage,which mainly consist of material con-ditions,the effectiveness of the equipment and facilities,the operation site,and the safety management as well as the emergency rescue situation reliability.Besides the above said components,there are also 20 2 nd-oroler indicators functioning for the risk assessment in LNG storage and transportation enterprises.And,secondly,this experiment has also included the LNG leakage data set by taking a few of LNG storage and transportation enterprises as positive samples and some others with leakage accidents as negative ones.And,thirdly,PCA can also be used to analyze the data set and sample scatter plots,which can be drawn to investigate the effect of the first 2 principal components on the classification results.Thus,through the analysis of the first 5 principal components,it can be found that under each principal component,the 2 nd-order indicators can be ranked according to their respective weights,while the degrees of their impacts are likely to extend to the frequency of the said accidents.Thus,finally,the random forest classification method can also be used to classify and calculate the weights of all the indicators for the classification effect.Hence,through comparison,the results of PCA can let us conclude that the impact of the 2 nd-order indicators in the safety management and emergency rescue on the occurrence and expansion of the LNG leakage accidents ought to be greater than the 2 nd-order indicators in the material condition,equipment and facilities and the operation maneuvarability.Thus,it can be seen that the given study tends to be able to provide useful suggestions for the enterprises concerned to improve their LNG leakage risk management and control,and technically support them to manage the LNG risk big data successfully.