| With the rapid increase of motor vehicle ownership and highway mileage in China,increasingly prominent traffic safety problems are also accompanied.It is the key content of traffic safety improvement work at the present stage to carry out reasonable traffic safety evaluation research and formulate targeted rectification measures.The traffic safety measurement evaluation method based on accident frequency modeling is an important research branch of the traffic safety evaluation system,and also an important work link to ensure the highway traffic safety.However,the current research method mainly has the following defects:(1)the accident frequency modeling includes risk factors needs to be extended range,most of the research according to the design characteristics of domestic highway traffic safety and traffic related indexes evaluation,but the accident involving people-car-road-environment various factors,one-sided to take one or two kinds of risk factors will inevitably affect the reliability of the traffic safety evaluation.(2)There is a lack of systematic research on the main intrinsic attributes of the accident data set(including "multizero characteristics","heterogeneity of safety effect",and "spatial-temporal correlation of safety effect"),which leads to doubts about the accuracy of the accident frequency model and the reliability of the safety effect measurement analysis.(3)Limited by the quality defects of the above data and models,there is a gap between the results of expressway measurement evaluation and the actual situation,or even completely contrary conclusions.Aiming at the above problems,this paper takes several mountainous expressways in Guangdong province as the research object,and constructs multi-dimensional accident data sets of conventional and tunnel sections including road design characteristics,traffic conditions,pavement performance and weather conditions.On this basis,a series of innovative traffic safety measurement models are established by introducing Lindley and Generalized exponential distribution to fit multiple zero-valued samples,using the method of parameter randomization to characterize the heterogeneity of safety effect,and introducing conditional autoregression and first-order hysteresis effect to consider the temporal and spatial correlation of safety effect.Based on the results of Bayesian estimation and quantification index of safety effect of the innovative models,the influencing factors of accidents in conventional expressway sections and tunnel sections are deeply analyzed,and corresponding improvement strategies are formulated.The research content and innovation results of this paper are as follows:(1)Establishment of high-quality data sets and statistical test of modeling variables.Fusion of the five typical mountainous area of guangdong province highway traffic accidents,road design characteristics and weather conditions,traffic conditions,road surface performance multidimensional data,determine the "quarter" as the time unit,with road type,type of plane curve,longitudinal slope degree and number of lanes for the space division index,preliminary divided into a series of sample set.In view of the samples divided,the optimal length range of modeled sections was determined from the statistical distribution characteristics of accidents and the function relation method,and the final sample set of conventional sections and tunnel sections was obtained by further integrating the irregular sections.Then,29 risk factors for accidents in conventional expressway sections and 24 risk factors for accidents in tunnel sections were selected,and the characterization indexes of each risk factor were matched to the corresponding samples to form two preliminary modeling data sets.On this basis,Pearson correlation coefficient,Klemm correlation coefficient and Intraconsistency group correlation coefficient were proposed to test the collinearity of all variables,and finally,the accident frequency modeling data set of expressway regular section(a total of 26 significant risk factors)and tunnel section(a total of 24 significant risk factors)was established.Finally,taking the conventional road section data set as an example,the statistical distribution characteristics and variation rules of accident frequency and modeling independent variables are analyzed in detail,and the spatial and temporal agglomeration of some modeling variables is preliminarily confirmed,which provides high-quality data support for subsequent modeling method research.(2)Study on the statistical distribution types and characteristics of accident frequency.Firstly,the logic of accident occurrence and its correlation with Bernoulli test are explained in detail,and the process of accident logic conforming to Bernoulli test is determined.Then on the basis of a large number of research,this paper expounds the process of the construction of the four types of classic accident frequency models and their applicable conditions with the Poisson model,Negative binomial(NB)model,Zero expansion Poisson(ZIP)model and zero expansion negative binomial(ZINB)model.The parameter estimation method based on Bayesian estimation method is determined,and Pearson chi-square statistics is used to measure the coincidence degree of accident frequency with its assumed distribution form.Vuong statistics were used to compare the goodness of fit between the zero-expansion model and the corresponding traditional model.The criterion of deviation information,mean absolute deviation and root mean square error were used as the test indexes of goodness of fit.Finally,the performance of the above four classical accident frequency models is compared from two aspects of logical accident interpretation ability and goodness of fit by using the established data set of conventional expressway sections.The results show that the best distribution form of accident frequency in this study is negative binomial distribution.Although ZINB model achieves good fitting effect,it lacks the ability to explain the logic of accident occurrence.(3)Study on accident frequency modeling method based on multi-zero characteristics.In this paper,we analyze the influence of multiple zero-valued sample attributes on model performance,and discuss the advantages of Lindley distribution and Generalized Exponential(GE)distribution in fitting data with multiple zero-valued attributes.The hierarchical models of NB distribution and Lindley distribution,NB distribution and GE distribution(NB-L model and NB-GE model respectively)were proposed.On this basis,the NB-L model and NB-GE model are applied to the metrological evaluation of traffic safety in conventional expressway sections,and compared with the classical NB model and ZINB model,the superiority of NB-L model and NB-GE model is verified.Finally,based on the Bayesian parameter estimation results of the optimal model NB-L model,the safety effect quantification index with elastic coefficient as continuous variable and marginal effect as discrete variable was proposed to quantitatively analyze the safety effect of each significant accident risk factor in conventional expressway sections.The results showed that a total of 20 risk factors detected by NB-L model were significantly correlated with accident frequency,among which the risk factors that were positively correlated included:Service area section,tunnel section,overpass section,middle belt-2m,middle belt-3m,hedge lane,climbing lane,curvature,slope length,bridge proportion,cement pavement,ln(QADT),Class 1 vehicle,Class 5 vehicle,light/moderate rain,SRI;Negatively correlated risk factors include slope,class Ⅱ vehicles,class Ⅳ vehicles,PCI,RQI,heavy/rainstorm,and wind.(4)Study on accident frequency modeling method based on safety effect heterogeneity.The source of heterogeneity of data safety effect and its influence on the performance of accident frequency model were analyzed,and the improvement idea of characterizing heterogeneity of data safety effect by parameter randomization was determined.On this basis,the negative binomial Lindley(α-RENB-L)model of class I random effects was established by setting the discrete parameters of NB distribution as random parameters,based on the NB-L model with multiple zero-valued attributes.The intercept term in the link function was set as random parameter to establish the negative binomial Lindley(￡-RENB)model of class Ⅱ random effect,and the regression coefficient of variables was set as random parameter to establish the negative binomial Lindley(RPNB-L)model of random parameter.Finally,the above three types of accident frequency innovation models,which take into account the multi-zero and heterogeneity of data,are applied to the safety effect analysis of conventional expressway sections,and the feasibility of the improved heterogeneity method of parameter randomization is verified,and eight random variables are tested:Middle zone3m,no climbing lane,slope,ln(QADT),Category 1 vehicle,Category 5 vehicle,light/moderate rain,heavy/heavy rain.On this basis,by comparing the changes of regression coefficients of significant variables and the distribution of random parameters between NB-L model(without considering heterogeneity)and RPNB-L model(with considering heterogeneity most comprehensively),the effects of heterogeneity of safety effect on safety effect of all significant risk factors were systematically analyzed.(5)Study on accident frequency modeling method based on spatial-temporal correlation of safety effect.The influence of spatial-temporal correlation theory and spatialtemporal correlation of safety effect on accident frequency modeling was analyzed in detail,and the necessity of considering spatial-temporal correlation in tunnel accident frequency modeling was determined by using Moran index spatial correlation test method.Based on the RPNB-L model with multiple zero-value and safety effect heterogeneity,a spatially correlated stochastic parameter negative binomial Lindley(SP-RPNB-L)model was established by embedding conditional autoregression-prior(CAR)into the link function.The time term with first-order hysteresis effect is embedded to form a time-dependent random parameter negative binomial Lindley(T-RPNB-L)model.The negative binomial Lindley(ST-RPNB-L)random parameter model of spatio-temporal interaction was formed by embedding the above spatial term,time term and spatio-temporal interaction term.On this basis,the effects of the established spatio-temporal correlation series model and RPNB-L model on the traffic safety measurement evaluation of the tunnel section are compared and analyzed,which verifies that st-RPNB-L model has the best goodness of fit.The results also show that 12 risk factors have a significant impact on tunnel accidents:Speed limit,Tunnel Length,tunnel entrance,curvature,Curve length,SRI,ln(QADT),Category 1 vehicle,Category 5 vehicle,light/moderate rain,heavy rain.Finally,according to the Bayesian parameter estimation results and safety effect quantification indexes of ST-RPNB-L model,the influence of 12 significant risk factors on highway tunnel traffic safety is analyzed in depth.(6)Study on accident cause and traffic safety improvement countermeasures based on quantitative analysis of safety effect.According to the parameter estimation results and safety effect evaluation indexes of a series of accident frequency innovation models,the importance degree of safety effect of each risk factor is ranked,and 7 accident risk sections and periods are summarized,and targeted traffic safety improvement measures are put forward.On this basis,taking the long downhill section of Kaiyang expressway as an example,the accident cause analysis and traffic safety improvement measures are studied.Finally,according to the results of the paper,the specific process of analysis method of expressway accident cause and improvement measures under the condition of all factors is put forward.The research results of this paper provide theoretical basis and technical support for analyzing the mechanism of expressway traffic accidents and predicting the development trend of traffic safety,which has important reference value for improving expressway traffic safety. |
PDF Full Text Request