Automatic detection of spatial safety clearances in substations based on 3D point cloud semantic segmentation

Abstract

Automated spatial-clearance control in substations requires optical perception that can recognize compact electrical assets, quantify geometric uncertainty, and issue traceable control decisions. This paper presents a dual-scale geometry-guided semantic segmentation and clearance estimation method (DGSCE) for registered LiDAR point clouds. Fine and coarse covariance neighborhoods characterize slender energized conductors, planar equipment surfaces, and local context; a class-balanced ExtraTrees ensemble produces point-wise posteriors, and confidence-gated graph repair suppresses isolated la- bels without globally smoothing boundaries. Density-connected components are associated by nearest-neighbor search, and range and voxelization uncertainties are combined in an explicit reserve to obtain a conservative clearance. A quality-gated supervisory interface then maps the conservative distance to CONTINUE, HOLD/INTERLOCK, or HUMAN REVIEW states for inspection robots and work-zone access systems. A physics-constrained benchmark containing 48 scenes and 126,096 labeled points was used for controlled validation. On the 12-scene test split, DGSCE achieved 98.80% overall accuracy, 97.72% mean intersection over union, and 98.84% macro-F1; the raw clearance mean absolute error was 13.3 mm. At an illustrative 1.20 m threshold, all six unsafe scenes generated correct HOLD/INTERLOCK commands and all six safe scenes generated correct CONTINUE commands. Severe range noise and point dropout rapidly degraded alarm reliability, confirming that automatic action must be disabled by a validated quality gate outside the operating envelope. These results establish a verifiable optical-sensing-to-supervisory-control pipeline under controlled conditions.