Abstract: In order to improve the predictive ability of the railway accident duration and the instruction level of railway traffic dispatching, the 2006-2015 China Railway Accident Report data was selected as the research object. Firstly, the FL-BiGRU classification model was used to classify the direct causes into 7 different categories according to the characteristics of the accident duration. Next, the Maximal information coefficient of statistical methods was used to make a correction analysis of the long-term indicators and select the first eight significant factors related to the duration as predictors. Finally, five machine learning algorithms were used to establish the accident duration prediction model to verify the rationality of using the correlation analysis to filter the factor set. The results show that the accuracy of the FL-BiGRU classification model reaches 94%, and the prediction index system constructed by the maximum information coefficient can significantly improve the accuracy of the prediction model. The prediction accuracy of the CHAID model, which can reach 79%, is better than other models. The CHAID model can be used to assist the dispatching work in practical work.

Key words: railway accident duration, railway accident reports, focalloss, bidirectional gated recurrent unity, maximum information coefficient