Research On Railway Accident Duration Prediction Method

With the rapid development of railway transportation and the improvement of automation,the ability of rapid response to various risks and accidents has become the reliable guarantee of transportation service quality.However,some uncontrollable factors such as fixed and mobile equipment failure,bad weather will lead to various kinds of accidents,affecting the normal operation of the railway plan and reducing passenger satisfaction.Therefore,estimating the duration of the accident in advance,providing information support for the rapid optimization of the dispatching system,and making the system resume normal operation as soon as possible,become the current research focus.Based on the text mining and feature selection,a prediction model was designed according to the distribution characteristics of railway accident duration to estimate the accident duration in this paper.The main contents of this paper are as follows:(1)Bidirectional Gated Recurrent Unit with Focal Loss(FL-BiGRU)was proposed for the unbalanced distribution of causation factors affecting railway accident duration to improve the recognition accuracy of small sample class by using Focus Loss(FL)function to train model and automatically adjust the weight of loss function.Experimental results showed that the proposed FL-BiGRU algorithm had higher classification performance than the traditional BiGRU algorithm,and can accurately classify small sample causative categories.(2)Based on the study of the characteristics of railway accidents in China,the characteristics of accident duration were summarized in this paper,including causation factors and external conditions.According to the characteristics of various and miscellaneous railway accident attributes,998 railway accident features were coded to construct the Railway Accident Duration Feature Set(RAD-FS).(3)A Hybrid Algorithm(HA?mRMR-MIC)combining the improved Filter Algorithm(mRMR based on MIC,mRMR-MIC)and Embedded Algorithm was designed to select the features of RAD-FS.The mRMR-MIC was used to arrange the feature sets in descending order,then combined with the Embedded Algorithm,the sorting results of mRMR-MIC were added up as the input of the classifier,and the classifier results were used to select the optimal k value of the predictive feature set to construct the Railway Accident Duration Prediction Feature Set(RADP-FS).(4)According to the characteristics of long tail distribution of railway accident duration,the Bi-level Prediction Model for Railway Accident Duation(BPM?RAD)combining classification and regression was proposed.The classification model was applied to identify the short-term accidents,and the regression model was applied to predict the duration of railway accidents in minutes.In addition,based on the characteristics of railway accident duration distribution,XGBoost algorithm with tree enhancement,loss function regularization,and adaptive learning rate was selected to construct classification and regression models.(5)An improved XGBoost with Bayesian Optimization(XGBoost?BO algorithm)was put forward to solve the problem that XGBoost model has too many parameters and too much searching space.Bayesian optimization algorithm was applied to optimize the parameters of XGBoost model,then XGBoost?BO classification prediction model and XGBoost?BO regression prediction model were constructed.Finally,the validity of XGBoost?BO algorithm was verified by comparing different tree models.