Model evaluation demonstrates accurate accident classification from construction hazard reports, suggesting viable automated risk monitoring for large-scale hydropower projects.
Key Points
To develop and evaluate an interpretable hybrid deep learning model capable of inferring potential accident categories from unstructured textual hazard reports in hydropower engineering construction.
Analyzed 9,606 Chinese-language hazard reports collected from the Baihetan Hydropower Station to classify text into 15 potential accident categories.
Engineered a hybrid architecture integrating RoBERTa for contextual semantic representation, a bidirectional long short-term memory (BiLSTM) network for sequential dependencies, and hierarchical attention for multi-level feature aggregation.
Implemented SHapley Additive exPlanations (SHAP) to interpret prediction outputs by highlighting critical words and phrases driving model classifications.
The proposed hybrid deep learning approach achieved a mean precision of 89.83%, recall of 88.12%, and an F1-score of 88.97%.
The model outperformed the strongest baseline architecture (RoBERTa + BiLSTM) by an absolute increase of 1.91 percentage points in F1-score.