Knowledge-Guided and Inspection-Data-Compensated Rapid Scene Reconstruction for Multi-Robot Maintenance of Overhead Contact Systems

Rapid generation of accurate three-dimensional (3D) maintenance scenes is a key prerequisite for multi-robot operation in overhead contact system (OCS) maintenance. Existing rule-driven modeling methods are efficient but cannot reflect in-service geometric deviations, whereas data-driven reconstruction methods are accurate but often too time-consuming for short maintenance windows. This paper proposes a knowledge-guided and inspection-data-compensated rapid scene reconstruction method for task-level robotic OCS maintenance. A pre-constructed PhyGeo-KG provides task-indexed scene templates, IFace assembly anchors, and physical constraints, while real inspection data are used to compensate local geometric deviations. The method restricts reconstruction to a two-span three-pole local region and instantiates the scene through three coupled levels: macro-level template retrieval and parametric layout generation, meso-level IFace-based affine assembly, and micro-level deviation compensation through sparse Jacobian mappings. To preserve minute-level efficiency, continuous wire geometry is corrected using a local parabolic approximation derived from the catenary model. Experiments on dropper replacement, cantilever-bolt tightening, and stagger adjustment scenarios show that the proposed method completes reconstruction in 1.65, 1.33, and 1.75 min, respectively. Ablation results demonstrate that the coupled method reduces contact-wire height and dropper-length deviations to ±5.2 mm and ±3.9 mm while preserving 100% component completeness and 98% assembly correctness. The 3D RMSE at ten key detection points is 8.7 mm, indicating sufficient fidelity for downstream multi-robot simulation and collision checking.