Abstract：In-depth research on semantic segmentation algorithms for substation point clouds is crucial for advancing the development of smart grids. Current semantic segmentation algorithms for substation point clouds rely predominantly on fully supervised learning, which requires extensive manual annotation of point cloud data. This makes the segmentation process time-intensive and costly. Weakly supervised methods have demonstrated significant advantages in reducing the human effort required for annotation, emerging as a prominent research focus in point cloud semantic segmentation. However, existing weakly supervised methods primarily adopt a single information propagation strategy, which struggles to accommodate the complexity and diversity of substation point cloud data structures. Moreover, these methods often fail to adequately capture local connectivity patterns in the point cloud, resulting in inefficient semantic feature propagation and diminished segmentation performance and accuracy. To address these challenges, this study proposes an improved PointNet-based algorithm named WSS-PointNet. The approach begins with constructing a multi-layer down-sampling structure, employing multiple sampling and grouping layers to process the input point cloud data and extract multi-scale feature representations. This multi-scale feature extraction effectively captures geometric and topological information of the point cloud at various scales. Subsequently, the PointNet structure is introduced to further extract regional features. The PointNet layer retains the multilayer perceptron (MLP) and max-pooling operations of the original PointNet architecture while removing the T-net structure to simplify the network and reduce computational complexity. A feature aggregation strategy is then used to integrate local features into a global feature representation. This process preserves local details in the point cloud while enhancing the model's ability to understand and recognize the overall scene by extracting global features. Through this method, the model effectively learns rich feature information from the point cloud data, providing coarse-grained semantic features for the subsequent network architecture. Following coarse-grained feature extraction, two semantic information embedding mechanisms are proposed: dilated semantic information embedding and immersive semantic information embedding. These mechanisms employ “inside-out” and “outside-in” information propagation strategies, respectively, to finely optimize point cloud features. Both embedding structures utilize graph convolutional neural networks (Graph-CNNs). The first GCN layer captures local connectivity patterns within the point cloud and shares effective information between the two network structures. In the second GCN layer, dilated and immersive embedding mechanisms enable morphological dilation and immersion, respectively, enriching the semantic features of the point cloud. To reduce overfitting during training and improve the model's generalization capability, three data augmentation techniques are introduced: pose transformation, point jittering, and attribute attention mechanisms. These methods effectively increase the diversity of training data. In this study, Avia LiDAR devices were used to collect point cloud data at substation sites, including the Xingtai substation in Hebei Province. The original dataset includes critical electrical facilities such as transformers, switchgear, towers, conductors, and monitoring equipment, as well as environmental elements like vegetation and buildings. After data augmentation, the total point cloud data volume reached 4 484 089 points, with 22% of the data generated through augmentation. The augmented dataset was further divided into training and testing sets in an 82 ratio. Experimental results demonstrate that WSS-PointNet improves the overall accuracy (OA) of substation point cloud segmentation by 10.3 percentage points, the mean intersection over union (mIoU) by 10.1 percentage points, and the mean accuracy (mAcc) by 10.5 percentage points, compared to its predecessor. Additionally, it reduces the annotation time required by 90%, achieving segmentation performance comparable to the best fully supervised algorithms. This model significantly reduces the time and cost associated with processing substation point cloud data while maintaining high segmentation accuracy.
裴少通, 孙海超, 胡晨龙, 王玮琦, 兰博. 基于WSS-Pointnet的变电站点云弱监督语义分割方法[J]. 电工技术学报, 2026, 41(1): 234-245.
Pei Shaotong, Sun Haichao, Hu Chenlong, Wang Weiqi, Lan Bo. A Weakly Supervised Semantic Segmentation Method for Substation Point Clouds Based on WSS-Pointnet. Transactions of China Electrotechnical Society, 2026, 41(1): 234-245.
[1] 贾惠彬, 武文瑞, 吴堃, 等. 基于异步整形机制的智能变电站通信队列调度策略[J]. 电工技术学报, 2024, 39(17): 5422-5433. Jia Huibin, Wu Wenrui, Wu Kun, et al.A communica-tion queue scheduling strategy for intelligent substations based on asynchronous shaping mechanism[J]. Trans-actions of China Electrotechnical Society, 2024, 39(17): 5422-5433. [2] 刘刚, 胡万君, 刘云鹏, 等. 降阶技术与监测点数据融合驱动的油浸式变压器绕组瞬态温升快速计算方法[J]. 电工技术学报, 2024, 39(19): 6162-6174. Liu Gang, Hu Wanjun, Liu Yunpeng, et al.A fast calculation method for transient temperature rise of oil immersed transformer windings driven by fusion of order reduction technology and monitoring point data[J]. Transactions of China Electrotechnical Society, 2024, 39(19): 6162-6174. [3] 雷蕾潇, 何怡刚, 姚其新, 等. 基于变权属性矩阵的变压器零样本故障诊断技术[J]. 电工技术学报, 2024, 39(20): 6577-6590. Lei Leixiao, He Yigang, Yao Qixin, et al.Zero-shot fault diagnosis technique of transformer based on weighted attribute matrix[J]. Transactions of China Electrotechnical Society, 2024, 39(20): 6577-6590. [4] 姜骞, 刘亚东, 严英杰, 等. 面向高变倍场景的变电站巡检机器人云台相机对准方法[J]. 中国电机工程学报, 2024, 44(8): 3337-3347. Jiang Qian, Liu Yadong, Yan Yingjie, et al.Substa- tion inspection robot PTZ camera alignment method for high zoom scenes[J]. Proceedings of the CSEE, 2024, 44(8): 3337-3347. [5] 普子恒, 陈志刚, 史星涛, 等. 考虑合成电场影响的换流站直流场巡检机器人路径优化[J]. 高压电器, 2023, 59(5): 178-184, 192. Pu Ziheng, Chen Zhigang, Shi Xingtao, et al.Path optimization of DC field patrol inspection robot in converter station considering the influence of synthetic electric field[J]. High Voltage Apparatus, 2023, 59(5): 178-184, 192. [6] 杨帆, 吴涛, 郝翰学, 等. 基于谱图理论的变压器区域大规模点云轻量化方法[J]. 电工技术学报, 2024, 39(23): 7528-7541. Yang Fan, Wu Tao, Hao Hanxue, et al.Large scale point cloud lightweight method for power transformer area based on spectral graph theory[J]. Transactions of China Electrotechnical Society, 2024, 39(23): 7528-7541. [7] 裴少通, 孙海超, 孙志周, 等. 基于DI-PointNet的变电站主设备点云高精度语义分割方法[J]. 电工技术学报, 2025, 40(9): 2917-2930. Pei Shaotong, Sun Haichao, Sun Zhizhou, et al.High-precision semantic segmentation of point clouds for primary equipment in substations based on DI-PointNet[J]. Transactions of China Electrotechnical Society, 2025, 40(9): 2917-2930. [8] Maturana D, Scherer S.VoxNet: a 3D Convolutional Neural Network for real-time object recognition[C]// 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Hamburg, Germany, 2015: 922-928. [9] Wu Zhirong, Song Shuran, Khosla A, et al.3D ShapeNets: a deep representation for volumetric shapes[C]//2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Boston, MA, USA, 2015: 1912-1920. [10] Boulch A, Guerry J, Le Saux B, et al.SnapNet: 3D point cloud semantic labeling with 2D deep segmentation networks[J]. Computers & Graphics, 2018, 71: 189-198. [11] Charles R Q, Hao Su, Mo Kaichun, et al.PointNet: deep learning on point sets for 3D classification and segmentation[C]//2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Honolulu, HI, 2017: 77-85. [12] Qi C R, Yi Li, Su Hao, et al.PointNet++: deep hierarchical feature learning on point sets in a metric space[C]//NIPS'17: Proceedings of the 31st Interna- tional Conference on Neural Information Processing Systems, Long Beach, USA, 2017: 5105-5114. [13] Gao Wei, Zhang Lixia.Semantic segmentation of substation site cloud based on seg-PointNet[J]. Journal of Advanced Computational Intelligence and Intelligent Informatics, 2022, 26(6): 1004-1012. [14] Yuan Qianjin, Chang Jing, Luo Yong, et al.Automatic cables segmentation from a substation device based on 3D point cloud[J]. Machine Vision and Applications, 2022, 34(1): 9. [15] Tao An, Duan Yueqi, Wei Yi, et al.SegGroup: seg-level supervision for 3D instance and semantic segmentation[J]. IEEE Transactions on Image Processing, 2022, 31: 4952-4965. [16] Shao Feifei, Luo Yawei, Liu Ping, et al.Active learning for point cloud semantic segmentation via spatial-structural diversity reasoning[C]//Proceedings of the 30th ACM International Conference on Multimedia, Lisbon, Portugal, 2022: 2575-2585. [17] Unal O, Dai Dengxin, Van Gool L.Scribble- supervised LiDAR semantic segmentation[C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), New Orleans, LA, USA, 2022: 2687-2697. [18] Wang Puzuo, Yao Wei.A new weakly supervised approach for ALS point cloud semantic segmentation[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2022, 188: 237-254. [19] Hu Qingyong, Yang Bo, Fang Guangchi, et al.SQN: weakly-supervised semantic segmentation of large- scale 3D point clouds[C]//17th European Conference, Computer Vision-ECCV 2022, Tel Aviv, Israel, 2022: 600-619. [20] Zhang Renrui, Wang Liuhui, Qiao Yu, et al.Learning 3D representations from 2D pre-trained models via image-to-point masked autoencoders[C]//2023 IEEE/ CVF Conference on Computer Vision and Pattern Recognition (CVPR), Vancouver, BC, Canada, 2023: 21769-21780. [21] Liu Minghua, Zhou Yin, Qi C R, et al.LESS: label-efficient semantic segmentation for LiDAR point clouds[C]//17th European Conference, Computer Vision-ECCV 2022, Tel Aviv, Israel, 2022: 70-89. [22] Liu Jinxian, Chen Ye, Ni Bingbing, et al.Joint global and dynamic pseudo labeling for semi-supervised point cloud sequence segmentation[J]. IEEE Trans- actions on Circuits and Systems for Video Technology, 2023, 33(10): 5679-5691. [23] Li Junfeng, Yan Fei, Liu Yisha, et al.Weakly supervised semantic segmentation for point clouds with self-purification on pseudo labels[C]//2023 13th International Conference on Information Science and Technology (ICIST), Cairo, Egypt, 2023: 100-107. [24] Xue Wenhao, Yang Yang, Li Lei, et al.Weakly supervised point cloud segmentation via deep morphological semantic information embedding[J]. CAAI Transactions on Intelligence Technology, 2024, 9(3): 695-708. [25] Su Yanfei, Cheng Ming, Yuan Zhimin, et al.Multistage scene-level constraints for large-scale point cloud weakly supervised semantic segmentation[J]. IEEE Transactions on Geoscience and Remote Sensing, 2023, 61: 1-18. [26] Zhang Yachao, Qu Yanyun, Xie Yuan, et al.Perturbed self-distillation: weakly supervised large- scale point cloud semantic segmentation[C]//2021 IEEE/CVF International Conference on Computer Vision (ICCV), Montreal, QC, Canada, 2021: 15500-15508. [27] Wang Yue, Sun Yongbin, Liu Ziwei, et al.Dynamic graph CNN for learning on point clouds[J]. ACM Transactions on Graphics, 2019, 38(5): 1-12. [28] Wang Xu, He Jingming, Ma Lin.Exploiting local and global structure for point cloud semantic segmenta- tion with contextual point representations[C]// NIPS’19: Proceedings of the 33rd International Conference on Neural Information Processing Systems, Vancouver, Canada, 2019: 4571-4581. [29] Wang Junjue, Zhong Yanfei, Zheng Zhuo, et al.RSNet: the search for remote sensing deep neural networks in recognition tasks[J]. IEEE Transactions on Geoscience and Remote Sensing, 2021, 59(3): 2520-2534. [30] Li Yangyan, Bu Rui, Sun Mingchao, et al.PointCNN: convolution on X-transformed points[C]//NIPS'18: Proceedings of the 32nd International Conference on Neural Information Processing Systems, Montreal, Canada, 2018: 828-838. [31] Thomas H, Qi C R, Deschaud J E, et al.KPConv: flexible and deformable convolution for point clouds[C]//2019 IEEE/CVF International Conference on Computer Vision (ICCV), Seoul, Korea, 2019: 6410-6419. [32] Fan Siqi, Dong Qiulei, Zhu Fenghua, et al.SCF-net: learning spatial contextual features for large-scale point cloud segmentation[C]//2021 IEEE/CVF Confe-rence on Computer Vision and Pattern Recognition (CVPR), Nashville, TN, USA, 2021: 14499-14508. [33] Su Yanfei, Liu Weiquan, Yuan Zhimin, et al.DLA-Net: learning dual local attention features for semantic segmentation of large-scale building facade point clouds[J]. Pattern Recognition, 2022, 123: 108372. [34] Hu Qingyong, Yang Bo, Xie Linhai, et al.RandLA-net: efficient semantic segmentation of large- scale point clouds[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, USA, 2020: 11105-11114. [35] Yin Chao, Yang Bo, Cheng J C P, et al. Label-efficient semantic segmentation of large-scale industrial point clouds using weakly supervised learning[J]. Automation in Construction, 2023, 148: 104757.
2026, Vol. 41 
