Research On The Traffic Accident Severity At Highway-Railroad Crossings Considering Class Imbalance

The safety situation of railway transportation in our country is getting better and better,but motor vehicle-train collision accidents at highway-railroad crossings still occur from time to time.Such accidents also seriously affect the safety of people’s lives and property.Therefore,the paper constructs an accident severity prediction model for level crossing traffic accidents,analyzes the potential relationship between each influencing factor and the severity of the accident,and based on this,puts forward reliable suggestions and measures for the safe operation of the crossing,so that provide the relevant theoretical basis and technical support for the safety management of level crossings.The paper is based on the 2014-2018 motor vehicle-train collision accident data set of the public level crossing in the Federal Railway Administration(FRA)open source database.Considering the class imbalance of traffic accident data,the paper builds prediction models under different sampling methods and compares the comprehensive performance of the classification algorithm under each sampling technique.First of all,the paper preliminarily selected 36 independent variables with reference to existing research literature,and explained the meaning and quantification standard of each variable in detail.Secondly,after data cleansing,missing value filling and feature engineering and other processing procedures,37 2-class independent variables are finally obtained.Then,on the basis of summarizing the existing research results,the thesis extends the K-means+SMOTE sampling technique to the class imbalance processing of the traffic accident data of the 3 levels of level crossings,and compares it with SMOTE,BorderlinesSMOTE sampling techniques.The comprehensive performance of the traffic accident severity prediction models are compared and the results show that the XGBoost classification algorithm under the K-means+SMOTE sampling technology has better prediction performance and better analysis capabilities for fatal accident data.Finally,the paper proposes KACOS sampling technology based on K-means clustering and Ant colony algorithm,and builds models for predicting the severity of level crossing accidents based on this sampling technology and the XGBoost classification algorithm.The results show that this sampling technology can further improve the comprehensive performance of XGBoost.In addition,the paper analyzes the positive and negative effects of the main factors on the severity of the accident based on the SHAP value,and provides reliable suggestions and measures for the safe operation of the level crossing based on the analysis results.