小样本条件下的施工安全巡查报告文本分析及风险智能判别OACHSSCD

To address the problems of limited sample size and imbalanced category distribution in construction safety inspection reports,this study proposes a BERT-based ensemble text classification method for automatic identification of risk warning levels.Back translation combined with Easy Data Augmentation(EDA)is employed to enrich text samples and alleviate the shortage of minority-class data.On this basis,multiple training subsets are generated through bootstrap sampling,and a Bagging ensemble with a majority voting strategy is adopted to improve prediction stability under small sample condition.In addition,Focal Loss is introduced to reduce the bias toward majority classes during mod-el training.A case study based on safety inspection report from a metro foundation pit project is con-ducted to validate the proposed method.The results indicate that the proposed approach effectively im-proves the classification performance of risk warning levels under small sample and imbalanced data condition,achieving an overall accuracy of 96.24％on the test set.The proposed method provides a feasible technical solution for efficient utilization of textual information in construction safety manage-ment.