AbstractAiming at the problem of low detection accuracy due to insufficient feature extraction of point cloud information in the 3D scene of substation by the existing methods, this article proposes a safety control method based on joint learning and PointCNN for the substation site. First, voxel filtering and statistical filtering methods are used to appropriately simplify the point cloud data, and the point cloud is smoothed using the moving least squares method. Second, the PointCNN's substation site safety control model is designed under the federal learning (FL) framework, and in the first stage, the cross self‐attention module (CSA) is introduced, as well as the point cloud features are extracted by the CSA module to help the model in the encoding process, and the point cloud features are extracted using the PointCNN by extracting the semantic features of each point, performing the prospective point separation and extracting the pre‐selected frames. In the second stage, features are extracted from each candidate box by point cloud region pooling operation, and the global semantic features obtained in the first stage are used for feature fusion in combination with other relevant point features in the enveloping box (local spatial point features, Li‐DAR distance‐depth information, laser reflection intensity etc.), and the classification header and regression header are used for the optimization of the 3D enveloping box and the prediction of the confidence level. Finally, the experiments demonstrate that the proposed method has the highest accuracy in the case of relatively low running time, and the APs for the medium difficulty scenario targeting pedestrians, transformers, bushings, and GIS equipment are 63.98%, 85.29%, 57.11%, and 69.23%, respectively. Therefore, the proposed method can effectively improve the safety control ability at substation sites.
Read full abstract