SCESNet: A Precise Semantic Segmentation Algorithm for Laser Point Clouds of Key Substation Equipment

ABSTRACT

The primary challenge for intelligent inspection robots in substations lies in accurately locating critical equipment such as transformers and busbars using LiDAR data. Existing point cloud segmentation methods for substations suffer from low accuracy. This paper proposes enhancements to the PointNet++ algorithm through network architecture optimization, improved optimizers and advanced data augmentation techniques, ultimately developing a specialized point cloud segmentation method for key substation components. Specifically, we introduce the inverted residual module and PointTransformerV2 module to enhance point cloud information interaction, strengthen feature extraction capability and mitigate gradient vanishing. The implementation of RAdam optimizer and targeted data augmentation further improves segmentation accuracy. We constructed a dedicated point cloud dataset for substation equipment and conducted comprehensive ablation studies and comparative experiments. Results demonstrate that our SCESNet outperforms other mainstream deep learning algorithms and existing substation segmentation methods in semantic segmentation of key equipment point clouds. The proposed network achieves high‐precision segmentation, providing technical support for digital operation and maintenance of substations and enabling autonomous inspection path planning for drones or robots.