With the continuous advancement of deep learning technology, research on point cloud segmentation algorithms has achieved significant results. Deep learning models, especially Convolutional Neural Networks (CNNs), have demonstrated excellent performance in processing image and video data. However, the unordered and irregular nature of point cloud data presents new challenges for the training of deep learning models. To overcome these difficulties, various deep learning-based point cloud processing methods have been proposed worldwide, among which the PointNet++ algorithm has attracted much attention due to its effective handling of the unordered nature of point cloud data. The method proposed in this thesis not only better captures local features but also improves the recognition ability of object details while maintaining computational efficiency. Through experimental verification on multiple public datasets, the proposed method has achieved certain performance improvements in the task of object part segmentation, mainly relying on the following two innovations. First, an improved training strategy set was proposed to enhance the performance of PointNet++. Secondly, more separable Multi-Layer Perception (MLP) units were introduced into the original PointNet++ model to achieve efficient model expansion. The model proposed in this thesis achieved an Overall Accuracy (OA) of 92.4% in classification on the ModelNet40 dataset, exceeding the benchmark model by 1.7%; the mean Intersection over Union (mIoU) on the S3DIS dataset reached 55.8%, representing a performance improvement of 2.6%.