基于决策树模型的煤矿安全事故严重程度分析与预测

为将数据挖掘技术应用于煤矿安全管理,通过对我国1999—2015年29 000多条煤矿安全事故数据的研究,系统分析了事故发生的区域、时间、类型和企业信息等因素对事故严重程度的影响及彼此之间的相关性。通过构建决策树分类模型,在给定事故相关信息的基础上,对事故严重程度进行分类预测;基于数据类别不平衡的特点,采用欠采样的抽样方法,同时利用梯度提升的组合分类器来提高分类精度。结果表明,采用的数据挖掘模型在预测不同严重程度的事故上均达到了较高精度。

Analysis and prediction of injury severity of coal mine accidents based on decision tree

The paper intends to present an analysis and predictive evaluation of the death-injury severity of the coal mine accidents based on the decision-making tree by using a study sample consisting of 29 000 coalmine accidents that have taken place during the period between 1999 to 2015 in the country in hoping to work out the dominant influential factors on such accidents.As a result,we have also found that the injury severities of such accidents may all involve the following regions,time periods and enterprise affiliations.To illustrate and describe all the problems involved more clearly and logically,we have developed a 3-rd decision classification model to predict the above said injury situations via the information and data concerned.So far as we know,the distribution of the injuries and death rates of some of the coalmine accidents have been obviously skewed for some of accidents leading to large number of injuries and deaths.Therefore,the accidents illustrated in this paper can better be divided into two categories:ordinary accidents and severe accidents.And,then,the mentioned accidents can also be classified by using the entropy as the isolated criteria of classification and regression tree (CART) model.What is more,the paper suggests to identify and classify two categories of accidents,that is,common accidents and severe accidents in terms of the performance evaluation and prediction,which has been adopted in the traditional evaluation metrics.Nevertheless,the traditional decision tree behaves inefficient in identifying the severe accidents due to the data inefficiency and imbalance.To meet such a severe challenge,it is necessary to apply the so-called under-sampling techniques derived from a dataset,which is easier to get balanced.Besides,we have also worked out an integrated classifier device known as the gradient boosting,which can enable us to enhance the classification accuracy,and,then,reweigh each base tree in accordance with the training errors.It is just on the above basis that the experiments we have done on the coalmine accidents can be found consistent with the 5-fold cross-over validation assessment as compared with the traditional classification models such as SVM,KNN,and neutral network.Thus,a better understanding of the influential factors and the characteristic features of the coalmine accidents can be achieved as a guide for safety management of such accidents in the future.