In substation robot inspection tasks, high-precision semantic segmentation of 3D point cloud data is one of the key technologies. Traditional point cloud semantic segmentation algorithms have certain limitations, making it difficult to handle complex 3D scenes. Deep learning methods have compensated for the shortcomings of traditional point cloud semantic segmentation algorithms and have become the main method for achieving point cloud semantic segmentation. However, existing point cloud segmentation methods for substations face issues such as high complexity, low accuracy, and gradient vanishing. To address these issues and achieve accurate segmentation of the main equipment point cloud in substations, this paper proposes a high-precision semantic segmentation method for substation main equipment point clouds based on DI-PointNet.

Firstly, on the basis of the PointNet++ network structure, a double-layer consecutive transformer (DLCTransformer) module is introduced. Key points are sampled through the DLCTransformer to enhance information interaction between point clouds and expand the effective receptive field. Secondly, a hierarchical key sampling strategy is adopted. The point cloud data is divided into the original dense point cloud space and a sparse point cloud space formed after farthest point sampling. These are then divided into multiple non-overlapping 3D windows, ultimately generating key values required for self-attention mechanism calculations, thereby reducing computational complexity, improving the model’s receptive field, and aggregating long-range context to achieve information interaction of substation-associated point clouds. Finally, an inverted residual module (InvResMLP) based on residual connections and inverted bottleneck design is added to the network. This enhances the model’s ability to extract complex structural features from substation point clouds while effectively reducing the gradient vanishing problem, making the algorithm more robust in handling complex substation scenarios and improving the accuracy of semantic segmentation of substation main equipment point clouds.

Additionally, to validate the segmentation effectiveness of the algorithm, this paper uses Avia LiDAR equipment to collect point cloud images of different devices at substations such as the Baobei substation in Baoding City. The original data includes transformers, switchgear, steel towers, insulators, maintenance equipment, and others (mainly vegetation and buildings). To simplify the point cloud data while filtering noise, the original input point cloud is first subjected to grid sampling with a grid size of 0.03 m. Data augmentation methods such as z-axis rotation, scaling, perturbation, and color reduction are employed. The initial window size is set to 0.12 m and is doubled after each down-sampling layer. The DI-PointNet is trained using the cross-entropy loss function and Adam optimizer with the following hyperparameters: initial learning rate of 0.001, batch size of 2, and 100 epochs. To ensure the reasonableness and accuracy of the experiments, the comparative algorithms used in this paper are trained using the same hardware platform, environment version, loss function, optimizer, hyperparameters, and training strategies as DI-PointNet.

Through ablation experiments and comparative analysis, the DI-PointNet algorithm proposed in this paper improves the overall accuracy (OA) value of substation point cloud segmentation by 3.4 percentage points compared to before the improvement, while reducing algorithm complexity. The proposed algorithm outperforms other mainstream deep learning algorithms and other point cloud segmentation algorithms in the power sector. The performance of this algorithm is close to the accuracy of manual segmentation and can achieve precise segmentation of substation point clouds.