The object and semantic segmentation tasks are widely researched in the computer vision domain. Segmentation in 3D point cloud faces lots of challenges due to its unstructured and permutation-invariance characteristics. For achieving the segmentation target in 3D point clouds, this paper proposed a Channel separate broad learning system (ChaSepBLS) with a shorter training time and more concise structure. The proposed ChaSepBLS adopt separate channel mapping to obtain rich point-wise geometric information in feature layers. Meanwhile, the k-nearest neighboring method (k-NN) is used to concatenate local features by multiple groups of downsampling scales in enhancement layers. By comparing the Moore-Penrose generalized inverse of the concatenating features from both feature and enhancement layers, the main training process of the ChaSepBLS is finished. When compared with the basic and several variants of broad learning networks, our ChaSepBLS algorithm achieve better segmentation performance on ShapeNet and S3DIS dataset.