Research On EW-LDA Based Railway Accident Contributor Analysis Method

The increasing degree of integration and automation of railway system equipment has led to an increase in the difficulty of human-computer interaction.Therefore,human and organizational factors have become an important cause of railway accidents.To prevent reoccurrence of similar accidents,it is necessary to analyze the cause of the accident and make scientific safety management plans.The railway accident report can be used to study the accident contributor and learn from the accident.Therefore,a text mining based railway accident causal analysis method was proposed.Text feature extraction algorithm was designed to extract the human and organizational factors in the accident,and quantitative analysis was conducted.The main research contents are as follows:(1)Extended Words Latent Dirichlet Allocation(EW-LDA)topic model was proposed to extract the cause of the accidents from the railway accident reports,and the importance of the word was determined by the TextRank algorithm.The semantic similarity between word distribution and document distribution was used to expand the topic terms.The experimental results showed that the EW-LDA topic model proposed in this paper can extract more features related to the cause of the accident compared with the traditional LDA model.(2)Based on the extracted accident cause features,a new Human Factors Analysis and Classification System-Railway Accidents(HFACS-RAs)methodology was structured to classify the causes in details.The preconditions of unsafe acts were further subdivided into the task conditions for unsafe acts,the environmental conditions and individual conditions for unsafe acts.These three categories were further divided into more specific factors according to the accident cause features.Considering the parent and subclass accident causes in the modified HFACS-RAs model,Support Vector Machine(SVM)was employed to complete the two-level classification of the accident cases and build a structured accident data set.(3)To find out the most critical factors of the accident,a quantitative model was designed from Bayesian Network(BN).The direction of the arc between the nodes was determined by the relationship between the upper and lower factors in the modified HFACS-RAs model.The Chi-square test combined with unconstraint 0/1 optimization was applied to obtain the optimal network structure.According to the characteristics of the accident data,the Logistic Regression model was adopted to estimate the conditional probability tables(CPTs);(4)The causality relationship of causes and consequences was obtained using the Bayesian Network inference,and the prime contributor to the accident can be figured out based on the sensitivity analysis of accident causal factors.