Research On Intersection Accident Severity Prediction Based On Hybrid Kernel Limit Learning Machine

With rapid economic development and motorization,traffic demand has increased dramatically,leading to a variety of traffic safety problems that have caused huge losses of life and property.As a key hub of the road traffic system,the intersection has a complex traffic flow and frequent traffic accidents.The starting point for solving traffic safety problems is to conduct scientific and reasonable analysis and prediction of accidents.For large samples and high dimensional traffic accident data,factor analysis model can play a good role in dimensionality reuction for accident influence factor analysis;in accident prediction,machine learning method has good applicability and prediction accuracy.Based on this,this thesis digs deeper into the intersection traffic accident data,and based on the factor analysis model,we study and analyze the intersection traffic accident influencing factors from four aspects: road,environment,vehicle,and driver,and understand their intrinsic mechanism on the accident.Optimization algorithms are combined with machine learning methods to build a model applicable to intersection traffic accident severity prediction,providing a more accurate and adaptable method for accident severity prediction.Firstly,this thesis takes the UK traffic accident data in 2019 as the research object,selects the accident data occurring at intersections,and uses the LOF algorithm to detect and clean the accident anomalous data points.Nineteen feature variables were selected based on four aspects: road,environment,vehicle,and driver,and the data were normalized.Next,an intersection accident factor analysis model was established,and six common factors were extracted to build up a three-tier intersection traffic accident impact factor assessment index system.The weighting model was improved to obtain the degree of importance of each influencing factor on the accident,in which lighting condition,speed limit,vehicle type,weather condition,age,intersection control,and driving status were the most significant factors influencing intersection accidents.Then,in order to improve the parameter finding ability of the Pelican optimization algorithm for the accident prediction model in this thesis,three improvements are proposed to the algorithm,generating the initial population based on the good point set theory,introducing the reverse differential evolution mechanism,and adaptive variation of the optimal individuals.The effectiveness and robustness of the improvement strategy proposed in this thesis are verified by comparing with other optimization algorithms with better performance.Finally,polynomial kernel function and Gaussian kernel function are selected as two kernel functions of the hybrid kernel limit learning machine,and the parameters of the hybrid kernel limit learning machine are optimized using the improved Pelican optimization algorithm,and an accident prediction model of the hybrid kernel limit learning machine based on the improved Pelican optimization algorithm is constructed to predict the severity of accidents at intersections.It is also compared with the prediction results of the extreme learning machine,random forest,support vector machine and probabilistic neural network models,and evaluated based on the accuracy rate,recall rate,F＿measure,Kappa coefficient,ROC curve and the new evaluation index constructed in this thesis,the accident severity prediction non-acceptance rate.The results show that the improved Pelican optimization algorithm hybrid kernel limit learning machine model performs the best.