Research on predicting building façade deterioration in winter cities using diffusion model

Abstract

Building façades, continuously exposed to natural environmental conditions, are susceptible to various forms of damage over time. Accurate prediction of such deterioration is essential for guiding the design, maintenance, and preventive conservation of buildings—a practice aligned with the concept of Restauro Preventivo (preventive conservation, Italian). Traditional approaches to damage prediction primarily rely on real-time monitoring or physics-based modeling. These traditional methods have a low degree of automation, rely on explicit parameter inputs, and require a large amount of labor and a long lead time. Recent advancements have demonstrated that Diffusion Models (DMs) are capable of generating high-resolution images with rich, diverse features, offering new potential for forecasting façade degradation. A dataset comprising multiple images of buildings on Rongshi Street in Harbin was constructed, and a suitable model architecture was identified through design-of-experiments methodologies. A customized training approach was developed, incorporating mesh-based control mechanisms and tailored dataset augmentation to enhance predictive accuracy. Both qualitative and quantitative analyses were conducted, with the refined model achieving an average Structural Similarity Index Measure (SSIM) score of 71.2 %. This indicates that the model adequately learns the complex building damage information and improves the decision-making process for re-repairing buildings after damage.