AutoDefect: Defect text classification in residential buildings using a multi-task channel attention network

Abstract

The sustainability of a building can be ensured through effective maintenance. Effective defect management, which is essential for maintaining the performance and longevity of buildings, requires regular defect inspections. Such inspections are expensive and time-consuming, traditionally taking the form of unstructured textual data. Classifying the collected data is complex, potentially leading to errors. A systematic classification system that considers a wide range of characteristics, including work type, defect location, defect element and defect type, is urgently needed. We propose a new automated defect text classification system (AutoDefect) based on a convolutional neural network (CNN) and natural language processing (NLP) using hierarchical two-stage encoders. A variant channel attention mechanism (the text squeeze-and-excitation block) is incorporated for one-dimensional CNN-based text modeling that extracts robust features for each encoder to improve classification performance. Testing the model on Korean textual defect data, AutoDefect outperformed three recent NLP models, BERT, ELECTRA and GPT-2, and was significantly more cost-effective, dramatically reducing the time required for defect management and minimizing human error.