Aiming at the problems of graph convolution-based point cloud classification models in extracting feature information from different semantic regions of the point cloud and efficiently utilizing aggregated high-dimensional features, a novel point cloud classification model is proposed, which combines dynamic adaptive graph convolution with multi-layer pooling. Specifically, residual structures is employed to construct deeper convolutions and learn feature information from different semantic regions of point pairs at different levels to generate dynamically adaptive adjusted convolution kernels that update the feature relationships of different point pairs, thus extracting more accurate local features. At the same time, the aggregated high-dimensional features are input into a multi-layer max pooling module to recover the discarded feature information from the first max pooling layer and obtain richer high-dimensional features to improve the accuracy of the classification model. The experimental results show that the proposed model achieves an overall accuracy of 93.3% and an average accuracy of 90.7% on the ModelNet40 dataset, which is significantly better than the current mainstream point cloud classification models, and has strong robustness.