Research On Intelligent Classification Method Of Accident Hidden Danger Based On Semantic Features

With the progress of information technology,work safety supervision gradually develops to the field of intelligence.As an important part of the overall deployment of safety supervision,the classification task of accident hidden danger is related to the analysis of the current regional supervision focus,as well as the optimization of the next stage of regional supervision deployment.Finally,it will realize the direction transformation from passive response to active supervision in advance.In the existing work mode,the safety officer takes the Beijing safety production accident hidden danger classification standard as the classification basis,and corresponds the accident hidden danger to the classification label one by one.However,similar to the scene of "one person with multiple crimes" in the judicial field,some potential accidents conform to the characteristics of multiple categories under the classification standard,and the existing way of one-to-one correspondence between potential accidents and labels fails to fully reflect the severity of potential accidents.In order to intuitively reflect the severity and relevance of accident hidden danger,it is necessary to realize multi label classification of single accident hidden danger from the perspective of text semantic feature extraction,and build an intelligent classification model for the field of safety production.This paper takes the data of production safety accidents in Beijing in 2018 as the research object,from the perspective of text semantic features,combined with deep learning model and text similarity calculation,constructs a multi label classification model of accident hidden danger,which truly reflects the number and name of the labels corresponding to the accident hidden danger,which is beneficial for law enforcement personnel to analyze key hidden danger and reasonably optimize work deployment.The main contents of this paper are as follows:(1)Research on multi label classification theory.This paper studies the basic theory of single label text classification and multi label text classification,studies the current classification methods and ideas,and analyzes the current application scenarios.(2)The intelligent rough classification model of accident hidden danger is constructed.For the four categories of potential accidents with less total tag combinations and high co-occurrence,the multi tag classification problem is transformed into multi classification problem by constructing tag combinations.Combined with the semantic features of the text,the hybrid text vector is constructed,and the attention mechanism is introduced into the basic model of textcnn to construct a rough classification model for four categories of potential accidents.(3)The intelligent fine classification model of accident hidden danger is constructed.For the 26 types of accident hidden dangers with large number of tag combinations and low co-occurrence degree of tags,the model structure of seq2seq is adopted,and the word and paragraph characteristics of the accident hidden danger text are combined to improve the hierarchical attention network model.The LSTM unit is used to replace the GRU unit,and the sentence level attention mechanism layer is introduced into the decoder.Through two improvements,the decoding effect of the model is improved.In this thesis,the multi label classification model is constructed for the 4 categories and 26 categories of potential accidents.In the task of intelligent classification of accident hidden danger,compared with the benchmark model,the two improved models improve the accuracy and recall rate of the model.The experimental results can prove that the improved hierarchical attention mechanism model can effectively improve the classification effect of multi label classification model.