A BERT model-based approach for causation analysis of electric power safety accidents

By extracting and classifying the information within accident reports, the process of accident analysis can be made more efficie nt and streamlined. However, the processing of accident reports solely relying on human resources is inefficient and susceptible to subjective judgment, mainly due to the complexity of accident records and varying standards. This paper proposes the BERT model, a deep learning approach based on accident causation theory, to analyze accident reports. Based on the scientific theo ry of accident causation analysis, a dataset of accident causation was created, and the BERT model was fine-tuned to classify the causes and severity levels of accidents. The model was fine-tuned over a range of 1-50 epochs, and it achieved the highest F1 value of 90.24% at the 17th epoch. Furthermore, the Precision and Recall values achieved were 93.29% and 92.71% respectively. Comparison between the BERT model and traditional machine learning models (SVM, Naive Bayes, Random Forest, and Logistic Regression) clearly demonstrates the superior performance of the BERT model. The results of the experiment indicate that the BERT model has better performance than traditional machine learning models in accident causation analysis. This stud y provides a intelligent and efficient method of accident causation analysis for electricity safety regulators.