Automatic Identification of Causal Factors from Fall-Related Accident Investigation Reports Using Machine Learning and Ensemble Learning Approaches

To enhance the performance of learning from past fall-related accidents, this study developed an innovative framework for automatically extracting every individual causal factor from accident investigation reports based upon the modified framework of the human factors analysis and classification system. Multiple techniques including the synthetic minority oversampling technique (SMOTE) algorithm for handling imbalanced data, soft voting with unequal weights for ensemble learning, and hyperparameter optimization were adopted to improve automatic identification of causal factors from unstructured text data. Experimental results denoted there were no classifiers with the best accuracy and F1 score unanimously for any of the 19 subcategories of causal factors. Therefore, one or more specific classifiers were preferred for predicting one specific causal factor with the best performance. Further comparative analyses between seven classifiers demonstrated that the ensemble learning model by the algorithm of soft voting (ELSV) could provide more stable predictions with low variance across different causal factors compared with individual machine learning models. It was suggested that the ELSV ought to be prioritized for collectively identifying all 19 causal factors. These findings are beneficial for substantial learning from past fall-related accidents with high efficiency and reliability, and valuable insights can be discerned and utilized for controlling the risk of fall-from-height at construction sites.