Abstract
Fire door defects in residential buildings negatively impact construction management by reducing fire safety effectiveness, increasing the likelihood of smoke and fire spreading, and consequently putting occupant safety at greater risk. To address this critical safety issue, this study proposes and evaluates five transformer-based text classification methods-BERT, RoBERTa, ALBERT, DistilBERT, and XLNet-for automated defect detection. These methods are optimized using both common and method-specific hyperparameters, resulting in 1,458 model variants evaluated through multiple metrics. Among these, the optimized RoBERTa achieves the highest performance, demonstrating F1 scores of 92.13% (frame gap), 87.29% (door closer adjustment), 78.17% (contamination), 82.89% (dent), 80.17% (scratch), 96.66% (sealing components), 98.43% (mechanical components), and 67.81% (others), yielding an average F1 score of 85.44%. Furthermore, RoBERTa significantly outperforms the other optimized 535 text classification models (ANN, SVM, DT, LR, 1D CNN, and LSTM). These results underscore the potential and effectiveness of transformer-based methods for safety management in real-world construction scenarios.