Research On Recognition Of Construction Accident Causation Paths And Patterns Analysis Using Text Mining Techniques

In spite of the rapid growth in construction industry and the application of digital managing system,casualties and losses caused by construction accidents are still increasing.In order to improve the safety situation in construction,it is imperative for companies to have a more developed accident managing and analyzing capability.Learning the causal mechanism and uncovering latent accident patterns from past experiences are effective means to prevent similar incidents from reoccurring.Therefore,this paper aims to build an accident process and analysis method in a data-driven approach using a real construction accident dataset.In this study,a recognition and mining analysis method of accident causation path was developed using text mining techniques.The occurrence of accident patterns and the law of management effectiveness were studied to provide reference for ensuring safety in construction practices.The main contents of this paper are as follows.(1)Based on the characteristics of domain,part of speech and grammar of construction accident reports,an adapted Text Rank extraction model integrating multiple features was designed to accurately identify causal information.The accuracy of the proposed model was verified by calculating the values of accuracy,recall,F1 and F2 score.Based on the results of keywords extraction,a compound term detection model using mutual information and left-right entropy was applied to recognize each element of accident causation paths.Compared with the manual labeling results,the recognition method proposed in this paper can accurately identify the direct cause,indirect cause,rectification measures,accident type and loss from accident reports.(2)According to the semantic relations and HFACS model,a classification framework of construction accident causes was built by accumulating the prior extraction results.Under the framework,411 construction reports were decomposed and coded to establish a digitalized construction accident dataset.Based on FP-Growth,a multilevel frequent pattern mining algorithm for accidents was proposed,which enhanced the data structure and analysis ability by adding hierarchical relations and mining process for different categories of accidents.A more comprehensive result for further analysis was obtained by mining frequent accident patterns(FAP),temporal frequent accident patterns(TFAP),frequent near-miss patterns(FNMP)and severe accident patterns(SAP).The experiment results showed that the multilevel frequent pattern mining algorithm had better performances than Apriori in respect of number,length and informativeness.(3)A frequent pattern analysis method based on Venn and comparative approaches was designed.Four categories of frequent accident paths were calculated and compared in the same and different time periods.During the same time period,Venn diagram was applied to disclose the law of rectification effectiveness and managing value of accident paths overlapping and transforming among different categories.In successive time periods,comparative analysis was performed to find the laws of temporal distribution of frequent accident paths and repeated failure of prevention and control measures.Near-miss incidents with potential risks escalating into accidents was also verified.Finally,targeted countermeasures of obtained key causes and patterns were proposed.