공동주택 방화문 하자리스트의 텍스트 마이닝을 이용한 하자유형 분류

Abstract

Although fire door defect classification results have been widely applied in various areas—such as scheduling maintenance workers and setting maintenance priorities—no previous studies have focused on fire door defects described in Korean text. This research addresses that gap by examining the performance of various machine learning and deep learning techniques for text-based detection of fire door defects. A dataset of 4,212 defect reports collected from 8,786 household units was annotated to include eight distinct defect types. Five traditional machine learning models and three deep learning models were trained using three different vectorization approaches. The analysis centered on evaluating classifier performance and the impact of vectorization methods. Among the traditional methods, LRwith N-grams achieved the highest accuracy (81.89%), while an LSTM-RNN with word embeddings performed best among the deep learning models (74.97%). Overall, the traditional models outperformed the deep learning models. For the test dataset, LRwith N-grams yielded an average F1-score of 82.18% across all defect categories. These findings highlight the robustness of traditional methods for this particular context, while also suggesting that further improvements could be realized through the use of broader datasets and more in-depth exploration of deep learning approaches.