Adaptive drone inspection strategy for bridge based on multi-level representation learning

Recently, UAV-based intelligent bridge inspection technology has seen widespread application. However, balancing efficiency and accuracy in inspection tasks remains a significant challenge. This study, driven by adaptivity, introduces a novel UAV inspection logic and a multi-level representation learning-based strategy, split into two stages: automatic rough inspection and adaptive fine detection. The strategy incrementally learns the spatial structure, component characteristics, and defect features of bridges. During the rough inspection, the UAV rapidly identifies component properties by integrating spatial clustering post-processing methods with point cloud semantic networks. Next, simulated field-of-view models and dimensionality reduction techniques compress the point cloud space, guiding the UAV to perform spatial inspections in planar geometry. In the fine detection stage, the hybrid Light-PVIT structure, optimized for spatial and channel dimensions, extracts defect features identified during the rough inspection. This prior information directs the UAV to conduct detailed inspections of defect areas. This strategy markedly enhances the efficiency and accuracy of bridge inspections, offering dependable technical support for bridge maintenance.