Diffusion-Based Generative Augmentation for Dataset Construction in Door-State Detection of Temporary Electrical Distribution Boxes on Construction Sites
Abstract Temporary electrical distribution boxes (TEDBs) are essential facilities for temporary power supply on construction sites, and their door states are closely related to equipment protection, electric shock prevention, rainwater intrusion control, and overall construction-site safety management. However, vision-based TEDB door-state detection is limited by the lack of publicly available datasets, difficulties in on-site data acquisition, and imbalanced sample distributions under complex working conditions. To address these challenges, this study constructs a task-specific Construction Site Temporary Distribution Box Door-State Detection Dataset (CSTDB-D) by integrating real construction-site images, laboratory-simulated images, online media images, and diffusion model-generated samples. The dataset supports the recognition of “closed” and “unclosed” TEDB door states. In addition, a generative augmentation pipeline based on Stable Diffusion with Low-Rank Adaptation (LoRA) is developed to improve the generation quality of TEDBs as long-tail industrial objects and enrich the training data distribution. Comparative experiments with multiple representative object detection models were conducted to evaluate the effectiveness of the original and augmented datasets. The results show that, after introducing the selected synthetic samples, the average Precision, Recall, mAP@0.5, and mAP@0.5:0.95 across all evaluated models increased from 0.929, 0.925, 0.967, and 0.764 to 0.939, 0.936, 0.973, and 0.781, respectively. These results demonstrate that the proposed generative augmentation strategy can increase sample diversity and improve TEDB door-state detection performance. This study provides a data foundation and methodological reference for intelligent inspection of temporary electrical equipment in construction environments.