Urban Facility Risk Assessment and Response Priority Decision-Making Based on Vision-Language Transformer

Abstract

This study addresses issues in urban management information systems to improve the intelligence level of urban facility risk identification and disposal priority decision. These issues include fragmented multi-source data, risk assessment relying on manual experience, and a lack of a quantitative basis for work order sorting. The study introduces Vision-Language Transformer (VLT) to construct an integrated multimodal framework for urban facility risk assessment and response priority decision-making. The framework takes road and municipal facility images, inspection records, and complaint work order texts as inputs. Through VLT, it fuses visual and linguistic features in a unified semantic space to obtain joint representations that can characterize both appearance defects and semantic hidden dangers simultaneously. Combined with business features such as location importance and historical failure frequency, it outputs risk scores and response priority rankings. Experiments select a public road damage dataset, dividing it into three subsets (A, B, C) corresponding to main road business districts, residential and public service facilities, and old areas. The optimized model of this study is compared with the Single-modal Vision Transformer-based Risk Assessment Framework (SVT-Risk) and the Early-Fusion Multimodal Transformer-based Risk Assessment Framework (EFMT-Risk). Results show that the accuracy of the optimized model in subsets A, B, and C is 0.91, 0.89, and 0.88; the mean average precision (MAP) in subsets A, B, and C reaches 0.87, 0.86, and 0.89, respectively. The overall performance is superior to the two comparison frameworks, indicating that the method can more accurately identify high-risk facilities and reasonably sort the disposal order; meanwhile, it can effectively alleviate the problems of inaccurate risk identification and unreasonable work order priority in traditional urban management information systems. Therefore, this study makes certain contributions to the field of urban facility risk assessment and intelligent urban management information system construction.