Research On Forecasting Highway Accident Duration Based On Accident Text Data

In recent years,with the rapid growth of the total mileage of China’s expressway,the situation of domestic expressway traffic safety is increasingly severe.On the one hand,traffic accidents can not only cause serious property losses,but also bring physical and mental trauma to the people involved.On the other hand,it can also lead to road traffic congestion,thus reducing the efficiency of the entire expressway network.In order to reduce the congestion and risk caused by traffic accidents,it is required that the expressway traffic management department should respond to traffic accidents in the first time,accelerate the speed of emergency rescue,and take reasonable and efficient emergency rescue strategy.And the accurate prediction of accident duration is the basis of carrying out these work.The traditional expressway traffic accident duration prediction model relies on specific accident characteristics,and is greatly affected by the source,quantity and quality of accident data.Moreover,a large amount of accident information is recorded in the free flow text in the unstructured form,which cannot be fully quantified.Therefore,it is difficult for traditional prediction models to meet the practical application requirements in terms of data mining or data feature extraction.Considering the rich accident information content contained in the text data,based on the natural language processing technology and using the multi-dimensional information transmitted by human language in the accident text corpus,the research on the prediction of expressway traffic accident duration is carried out.(1)Firstly,an accident text data set with a complete time label is constructed,which is mainly composed of two parts: one is the accident ledger data of Sichuan expressway;the other is the data of 4,808 accident microblogs published in the microblog account of "Sichuan Expressway",which is obtained by using the web crawler technology.Secondly,considering the delay of the accident detection stage and the accident response stage,the research scope of the accident duration in this paper is limited to include the accident clearance stage and the traffic recovery stage according to the relevant research and accident processing workflow.Finally,to a greater degree to meet travelers travel demand,reduce the influence of the outliers,and text of the collected accidents data preprocessing,and meet the requirements of the research data,a total of 4334,and verify that two kinds of source data word vector and tag the same characteristics,probability distribution,can be used as training set later model.(2)Based on the analysis of the characteristics of the accident text data,the specific segmentation word list and the stop word list for the prediction of the accident duration are constructed,and the TI-W2 Vec text vector model is established based on the Word2 Vec model and the TF-IDF model,so as to better highlight the differences among words.Secondly,considering the continuous time sequence of the accident duration,the TW-Fisher clustering model was established by referring to the idea of Fisher clustering algorithm to minimize the difference between the features of similar texts and maximize the difference between the features of different categories of texts.The ordered regression problem was transformed into an ordered classification problem.The performance of different classification models was compared and analyzed.The results show that the TW-Text RNN model has the highest accuracy,and the accuracy of TW-Text RNN model is 10% higher than that of Text RNN model.(3)Considering that there are a large number of unlabeled accident text data in the network information,and it is a costly and time-consuming arduous task to label new samples manually,in order to enrich the sample size of the training set and improve the generalization ability of the model,based on the TW-Text RNN model,A prediction model of network information accident duration migration based on TW-BiLSTM was established.By keeping the weight of the word vector layer unchanged,the bidirectional long and short memory layer was fine-tuned on a small number of labeled data sets in the target domain,so as to achieve the purpose of transfer learning.The results show that the accuracy of TW-BiLSTM model with transfer learning strategy is6% higher than that of TW-BiLSTM model without transfer learning strategy.