Graph neural network-driven text classification for fire-door defect inspection in pre-completion construction

Area of Science:

• Artificial Intelligence
• Computer Science
• Structural Engineering

Background:

• Defective fire doors compromise building safety, accelerating fire and smoke spread.
• Automatic identification of fire door defects is crucial for timely maintenance and resident safety.

Purpose of the Study:

• To develop and evaluate Graph Neural Network (GNN)-based text classification models for automatic fire door defect identification.
• To compare the performance of four GNN models (TextGCN, TextING, TensorGCN, BERT-GCN) with optimized hyperparameters.

Main Methods:

• Systematic hyperparameter optimization was performed for four GNN models.
• A comprehensive evaluation of 1008 model variants was conducted using multiple performance metrics.
• The models were trained and tested on a dataset of fire door defect descriptions.

Main Results:

• The optimized BERT-GCN model demonstrated superior performance in identifying fire door defects.
• BERT-GCN achieved high F1 scores across various defect categories, including frame gap (91.28%) and door closer adjustment (90.52%).
• Overall, BERT-GCN achieved an average F1 score of 85.46%, outperforming 2,430 other evaluated text classification models.

Conclusions:

• GNN-based approaches, particularly BERT-GCN, show significant potential for enhancing safety management in construction.
• The study validates the effectiveness of automated text classification for detecting critical fire door defects.
• This research contributes to improving fire safety protocols through advanced AI techniques.