Study On The Relationship Between Traffic Flow State And Crash Pattern On Freeway

With the rapid development of freeway in China,the development of national economy is promoted rapidly,and the process of regional industrialization and urbanization is accelerated.However,the problem of highway traffic safety is increasingly serious and become the main growth source of road traffic accidents,which not only seriously threatens the safety of people's lives and property,but also becomes the main cause of traffic congestion.In recent years,with the continuous improvement and application of advanced infrastructure such as the freeway dynamic traffic control system,it is possible to obtain massive and high-precision real-time traffic flow data.In order to improve freeway traffic safety,and to provide theoretical basis and technical support for the traffic management departments to formulate more targeted dynamic traffic safety management strategies and optimization measures,scholars have gradually enhanced the research on freeway real-time crash risk prediction and other related traffic safety.However,the existing highway traffic safety research still has the following limitations:(1)Traffic flow will show different dynamic features in different macroscopic traffic flow status,which is specifically reflected in the differences among traffic flow parameters,while drivers will show different driving behaviors under the joint action of different traffic flow parameters.Therefore,from the macroscopic phenomenon to the microscopic behavior,it can be found that different macroscopic traffic flow status have different safety features,while the existing research lacks the consideration of the traffic safety features of the traffic flow status;(2)The various patterns of crashes in different traffic flow status has different crash precursor and crash mechanism.But the existing dynamic traffic safety control system always overlook the differences above,and is lack of the attention to the various crash patterns in different traffic flow status.Therefore,it is difficult to be better targeted in the process of crash risk prevention,and the prediction accuracy and efficiency of crash risk prevention is reduced;(3)The previous real-time accident risk prediction model only focused on the prediction sensitivity,specificity and other indicators,but ignored the prediction cost.There is a lack of comprehensive consideration in the predictability analysis process.In addition,the predictability of different crash patterns in different traffic flow status is also different.Based on the National Natural Science Foundation of China(No.51322810)"traffic safety design and evaluation",this paper systematically studied the crash patterns and risk models under the typical traffic flow state of freeway.The main research work are as follows:(1)Analysis on Occurrence Mechanism and Precursor features of Crash Patterns under the Classical Macroscopic Traffic ConditionBased on the six levels of service and three phases traffic flow theory,traffic status were divided.Combined with the distribution of typical dangerous traffic flow variables,the traffic flow features of each crash pattern under different macroscopic traffic flow status were analyzed by radar map.The Bayesian Logistic Regression Model was used to quantitatively analyze the crash risk of different crash patterns in different traffic flow status.Random forest method was used to explore the occurrence mechanism of various crash patterns under different macroscopic traffic flow conditions.Based on different traffic flow status,the precursory features of each crash pattern were analyzed by establishing a Bayesian Random Parameter Sequence Logistic Regression Model.(2)Analysis on the Occurrence Mechanism and Precursor features of each Crash Pattern based on Traffic Safety StatusFrom the perspective of crash collision type and severity,the correlation analysis method was used to screen the dangerous traffic flow variables and obtain the traffic variable combination without correlation.By establishing a Bayesian Multinomial Logistic Regression Model to analyze the precursor features of each crash pattern,the extracted precursor features were ensured to have a significant correlation with the risk of each type of crash,and the variables of dangerous traffic flow were further screened.Using the k-value clustering analysis method,the classification number was reasonably determined by the Calinski/Harabasz pseudof index,and the traffic safety status was divided based on the crash collision type and severity.Radar map was used to analyze the distribution features of dangerous traffic flow factors of different crasn patterns in different traffic safety status,and a Bayesian Logistic Regression Model was established to analyze the correlation between different traffic safety status and crash collision type and severity.The random forest method is used to study the mechanism of various crash patterns in different traffic safety status,and the relationship between each traffic safety state and classical macroscopic traffic state was studied by using the nonlinear canonical correlation analysis method.Based on the above traffic safety state division,the precursory features of each crash pattern were studied by establishing the Bayesian Random Parameter Sequence Logistic Regression Model.(3)Predictability Analysis on Real-time Crash Risk Prediction Models of Different Crash PatternsThe Traditional Bayesian Logistic Regression Model was used to establish the real-time accident risk prediction model of each crash pattern under different traffic flow status.The prediction effect of each crash pattern under different traffic flow status was compared and analyzed from the perspective of AUC mean based on traffic flow state and AUC mean based on crash sample.The predictability of each crash pattern under different traffic flow status was analyzed by using two indexes of sensitivity and the number of prediction measures.Finally,from the three perspectives of the dangerous traffic flow state,the prediction precision mean based on the crash sample and the prediction precision mean based on the crash risk of the traffic flow state,the influence of different traffic state dividing methods on the predictability results of each crash pattern were analyzed.(4)Real-time Crash Risk Prediction Models of Different Crash Patterns based on Deep LearningBased on the description of the training process by the deep learning-based accident risk model,the number of iterations of different deep learning models was determined by setting the basic parameters and combining the relationship between the loss value of the model and the number of iterations.Based on Short and Long-term Memory Neural Network,support vector machine(SVM)and Bayesian Logistic Regression Model,comparative analysis of three kinds methods of real-time risk prediction model under different traffic flow status of each crash patterns prediction effect,and connecting with the prediction accuracy and prediction cost.Finally,comparative analysis of the predictability differences with different real-time crash risk prediction methods under different traffic flow status of the crash pattern.(5)The Portability Verification and Enhancement Method of Domestic Freeway Crash Risk Prediction ModelBased on the American freewaydata,a real-time crash risk prediction model was established by the traditional Bayesian Logistic Regression model.The Bayesian Updating method was used to set the prior probability density distribution of the real-time crash risk prediction model in China according to the parameters of the real-time crash risk prediction model of American freeway,so as to obtain the bayesian updated real-time crash risk prediction model in China.From three aspects of prediction effect,prediction accuracy and prediction cost,the portability of the real-time crash risk prediction model in China before and after bayes updating was verified.