Underground cable shaft has the characteristics of large scale,wide range and complex spatial distribution.In order to improve the efficiency of underground cable shaft screening and ensure the safe and reliable operation of underground cable,this paper proposes a novel PointNet++model based on broad learning system(BLS)and encoder- decoder for cable shaft point cloud semantic segmentation,termed as B-PointNet++.Firstly,in order to improve the feature learning ability and efficiency of PointNet++for solving large-scale point cloud data,a PointNet++Encoder- Decoder model is proposed.Meanwhile,BLS algorithm is introduced into the PointNet to replace the multilayer perceptron(MLP)and give full play to the efficiency of BLS randomization learning.Secondly,the point cloud data of underground cable shaft in Xiong'an were collected and the data set required by real semantic label was added to the model training.Finally,compared with the existing methods,the results show that B-pointnet ++has higher precision,recall,intersectionover union and F1 values compared with PointNet and Pointnet +十，it is beneficial to multi-objective segmentation of underground cable shaft scenes,and has great application potential.