Coverage-path-planned autonomous UAV inspection for deep learning-based corrosion grading of steel railway bridges

Information from the abstract

Steel railway bridges require repeatable inspection methods capable of resolving spatially heterogeneous corrosion while reducing dependence on difficult-access surveys and continuous manual flight control. This study presents RustScan-DL/UAV, an AI-enabled framework coupling pilot-supervised autonomous UAV coverage-path planning with zone-level deep-learning corrosion grading. The planner generates geometrically calibrated boustrophedon paths, sequences bridge elements by a distance-minimising tour, and checks battery feasibility. Each image zone is graded using a fine-tuned ResNet-50 fused with hue-saturation-value oxidation scoring, producing five operational grades derived from ISO 4628-3:2024 and an exploratory photographic corrosion-severity proxy calibrated on a limited ultrasonic dataset. Validation used BRIDGE-760, a corrosion-grading benchmark of 760 images (12,160 zones) comprising three external corrosion datasets and a 106-image Chumphon steel railway bridge UAV campaign. RustScan-DL/UAV achieved 93.7% zone-level accuracy and F1=0.921 for critical-zone detection; HSV fusion improved accuracy by 2.5 percentage points over the CNN-only baseline. ResNet-50 is retained as the study-optimal backbone because it is the only candidate with complete BRIDGE-760, fusion, critical-zone, INT8 onboard, and closed-loop validation; lighter modern backbones are treated as prospective alternatives pending equivalent controlled tests. In 14 pilot-supervised flights, the planner achieved 98.4% predicted coverage and reduced flight distance by 31.2% relative to a naive sweep. The available flight logs support a 79% positive predictive value for triggered re-inspections but do not retain complete adjudication of non-triggered regions; consequently, a flight-level false-negative rate cannot be estimated from the archived campaign data and no safety-completeness claim is made. The resulting zone-level maps may support selection of locations for confirmatory inspection, while ultrasonic gauging remains necessary for structural assessment.

Why this record is monitored

This record has an Impact Signal of 72/100 based on recency, source, collaboration, and bibliographic signals. It prioritizes monitoring and is not a judgment of research quality.

Related topics: Infrastructure Maintenance and Monitoring · Concrete Corrosion and Durability · Non-Destructive Testing Techniques