Abstracts:Abstract Crack detection is crucial for ensuring the durability, safety, and structural integrity of civil infrastructure. Traditionally, this task involves manual inspections and crack width measurements using a crack width comparator gauge. However, these methods are time-consuming, subject to subjective judgment, and prone to errors in spatial measurement. While automatic crack detection algorithms have been developed, most focus solely on a single issue using deep learning techniques. Comprehensive models that integrate crack segmentation, width estimation, and propagation assessment—essential for thorough structural evaluation—are still lacking. This paper presents the concrete health monitoring (CHM) system, a novel deep learning framework designed to enhance crack segmentation, width estimation, and propagation modeling in real-world scenarios. The CHM system employs a multimodal approach to concurrently perform these tasks. For crack segmentation, it introduces two innovations: multisource visual fusion (MVF) and attention-based hierarchical objective refinement (AHOR), addressing current methodological limitations. Width estimation is facilitated by a vision transformer regression model, and crack propagation is modeled using Paris’ law that correlates the crack growth rate with the stress intensity factor. Our results show that CHM achieves superior performance on a benchmark data set, with an accuracy of 87.26%, an intersection over union (IoU) score of 80.76%, and a recall rate of 84.51% for crack segmentation. For width estimation, it achieves a root mean squared error of 19.764. These outcomes affirm CHM’s efficacy in real-time infrastructure safety management.