Research On The Classification Of Hidden Danger Of Building Accidents Based On BERT Model

With the development of smart construction projects,the demand for smart management is increasing.Therefore,how to construct an information system of managing hidden dangers for construction enterprises has become a popular research topic,which is also a pain point of many enterprises.Safety inspection notification is the most popular method for detecting potential dangers in construction projects,recording the changes of hidden dangers in the whole process of the project under construction.However,it requires a lot of time and manpower to classify the notification manually,and it also requires personnel to acquire related knowledge.Therefore,it is significant to realize intelligent identification of potential dangers and improve the efficiency of a construction accident detection.To realize the automatic classification and identification of hidden danger and improve the practicability and applicability of safety inspection notification,this paper proposed a Bert improvement and compression model for the classification of hidden dangers in construction and established an automatic hidden danger identification interactive system based on the Tkinter.This paper will cover the following aspects:(1)The safety inspection notices in the corporate management system of a construction company from January 2013 to August 2020 were sorted out,a total of612 safety inspection notices were obtained,and data cleaning and denoising were performed.The manual preprocessing operation removed special characters,useless information,and full-width mixing and other corpus noises,and then performed corpus standards through the two-way exchange of data annotations,and finally sorted out16,033 hidden danger text data of construction accidents containing 12 hidden danger category labels set,which were used for subsequent model performance verification.(2)Three sets comparison model experiments were designed by utilizing Word2vec and Bert models as text representation experimental models,and continued to connect the same classification model to explore the semantic expression ability of the Bert model on the test corpus.(3)To address the existing problems in the experimental corpus,the author proposed a corresponding optimization scheme.To quantify the significant differences of the same terms among different hidden categories,this article used a TF-IDF-based term multi-category weighting scheme.For example,when a corpus consisting of n categories was given,each term would be assigned n weights,and each term was obtained by applying n category weight vectors to the word embedding matrix when it was input to the Bert model to improve the word embedding process of the Bert model.Aiming at the problem of unbalanced data distribution among hidden danger categories in the hidden danger corpus,the text applied a genetic algorithm to optimize the multi-category weightα_tof the Focal loss function.As a result,α_tfinally achieved the optimal value among 12-catogories under the continuous supervision and training of the custom fitness function.The category weight replaced the cross-entropy loss function in the Bert model that was not suitable for unbalanced data sets,thereby reducing the impact of unbalanced data distribution on the classification performance of the model in multi-text classification tasks.(4)For the Bert model,a model compression method based on model replacement—Bert-of-theseus was used,and the Bert-of-theseus compression effect was performed in sequence from four levels:different module replacement,different replacement rate,module replacement strategy,and teacher model layer number.Comparative analysis was conducted to explore the optimal compression effect to achieve the purpose of simplifying the model.(5)Based on the Ktinter framework,this article elaborated on the overall goals,functional requirements,and implementation process.According to the actual needs of construction companies,a construction accident hidden danger recognition system was constructed,the completed framework was visually displayed and finally passed the test.The corpus verification had completed the basic functions of the recognition system for hidden dangers.