Machine learning study demonstrates high-accuracy crack detection and concrete strength prediction in civil infrastructure, indicating the feasibility of unified condition assessment.
Key Points
To develop and evaluate a unified infrastructure condition-assessment framework that combines automated visual defect detection, material property prediction, and analytical evaluation within a single workflow.
Developed the Hierarchical Edge-Aware Residual Network (HEAR-Net) equipped with an Edge-Gated Activation mechanism to detect cracks under heterogeneous surface conditions.
Implemented a regression-based model to estimate concrete compressive strength from physicochemical composition parameters.
Integrated visual inspection and material property predictions into an analytical framework for holistic infrastructure condition assessment.
HEAR-Net achieved an accuracy of 99.90%, precision of 99.90%, recall of 99.90%, specificity of 99.90%, F1-score of 99.90%, and an ROC-AUC of 0.999993 on an independent test set.
The material property regression model predicted concrete compressive strength with an R² of 0.84 on the evaluated dataset.