Aiming at the difficulties and challenges caused by the irregularity, disorder and sparsity to point cloud analysis, a point cloud analysis method that combines local information extraction and global feature reasoning is proposed. First, in order to group local points more effectively, the structure-aware K nearest neighbor (KNN) is used to search for local neighborhood points. Secondly, a feature negative feedback convolution module is improved based on edge convolution to extract more accurate local features in the mapped high-dimensional space. In addition, a global semantic reasoning module based on the attention mechanism is designed to avoid potential information redundancy by emphasizing the grouping point of different regions, so as to obtain point cloud features more comprehensively. Through tested on the public point cloud data sets ModelNet40 and ShapeNet, the overall classification accuracy and overall mean intersection over union (mIou) of the proposed method reach 93.8% and 86.4%, respectively. Quantitative evaluation indicators and qualitative visualization experiments prove the accuracy and robustness of the proposed method.