Abstract: Existing point cloud deep learning methods were able to process point cloud data from fixed viewpoints. However, in practical applications, changes in object orientation affected the point cloud description by rotational transformations, thereby reducing recognition accuracy. To address this issue, a hierarchical rotation-invariant geometric structure representation learning method was proposed. First, point cloud samples were modeled using triangulated local geometric structures. A triangular surface was constructed within each neighborhood, and rotation-invariant features describing the geometric relationship between Euclidean space and the tangent plane were extracted. These extracted rotation-invariant features were then expressed through convolutional operators. Self-attention-enhanced convolutions were employed to aggregate local neighborhood structures, achieving adaptive fusion of local and global information. This further refined the rotation-invariant features and improved their global consistency. Finally, a hierarchical inverted residual multilayer perceptron module was introduced. Through multi-level nonlinear mapping and progressive channel expansion, shallow geometric features were hierarchically fused with deep semantic features, enhancing the high-level expressiveness and discriminative power of rotation-invariant features, and improving the ability to express and distinguish complex spatial structures and diverse rotations. The proposed method achieved an Overall Accuracy (OA) of 93.9% on the ModelNet40 dataset, an Overall Accuracy (OA) of 87.8% on the ScanObjectNN dataset, and achieved a mean Intersection over Union(mIoU) of 82.3% in the segmentation task of the ShapeNet dataset. The experimental results show that the proposed method demonstrates strong classification and segmentation performance, maintains rotation-invariance, and exhibits excellent robustness and generalization capabilities.