AI-Assisted Hybrid Deep Learning Framework for Early Micro-Crack Detection and Durability Monitoring in Sustainable Concrete

Abstract

Micro-crack initiation represents the earliest measurable indicator of durability degradation in concrete infrastructure and significantly influences long-term service life and structural reliability. Detecting sub-millimeter cracks (<0.3 mm), especially in sustainable and self-healing concrete with supplementary cementitious materials, remains challenging due to heterogeneous surface textures and environmental noise. Conventional convolutional neural networks frequently struggle to distinguish fine crack morphology from visually similar artefacts, such as scratches and aggregate boundaries. This study introduces a hybrid YOLO–Variational Quantum Classifier (VQC) framework for enhanced micro-crack detection and refinement. The YOLO stage enables real-time crack localization to maintain high recall, while extracted features are mapped into an expanded Hilbert space using angle encoding within an 8-qubit quantum circuit, allowing nonlinear geometric separation without explicit polynomial feature expansion. Entanglement operations capture higher-order inter-feature correlations, and variational parameters are optimized through gradient-based learning to construct adaptive nonlinear decision boundaries. Experimental evaluation on 8,500 annotated concrete surface images demonstrated that the proposed architecture achieved Precision = 0.94, Recall = 0.95, and F1-score = 0.95, outperforming CNN (0.84), ResNet-50 (0.87), standalone YOLO (0.92), and YOLO with classical refinement (0.93). The hybrid model reduced false positives while maintaining detection sensitivity. This framework supports predictive durability assessment, service-life modeling, and lifecycle-oriented management of sustainable concrete infrastructure, contributing to AI-driven innovation in concrete technology and construction material engineering.