Research on Point Cloud Segmentation Method Based on Local and Global Feature Extraction of Electricity Equipment

Intelligent inspection has become an important trend in the development of substation operation and maintenance technology, and fast and accurate point cloud segmentation of the large amount of point cloud data collected in the process of intelligent inspection is an important basis for effective fault localization and risk early warning of electricity equipment. However, the existing point cloud segmentation methods have problems such as messy spatial information, low classification accuracy and unstable model. Therefore, a new point cloud component segmentation method is proposed by combining the local and global information characteristics of the point cloud and applied to the diagnosis of electricity equipment. Firstly, apply dynamic-k neighbor retrieval for each point to facilitate subsequent local feature extraction. Then, the input point cloud data is processed by graph convolution to capture the local characteristics of the data; then, the global information is extracted by cyclic use of a set abstraction layer, and the two information features are fused and the model is trained to obtain a more powerful and efficient point cloud component segmentation model. Finally, experimental validation was carried out using point cloud data of lightning arrester devices in electricity equipment. The experimental results show that the accuracy and average mIoU of the part segmentation results of this method reaches more than 90%. With 2%-4% improvement and 4-10 times stability improvement compared to traditional point cloud processing methods, this method has better stability and accuracy in point cloud part segmentation tasks. This method not only provides an effective method for the segmentation of 3D point cloud data, but also provides strong technical support for the subsequent fault diagnosis and detection work.