<p>Existing 3D point cloud analysis methods prioritize translation and permutation invariance but generally lack systematic rotation-invariant encoding, degrading their performance on arbitrarily rotated 3D shapes. In this paper, we propose RiDBNet, an innovative rotation-invariant dual-branch network that incorporates complementary global and local rotation-invariant information through novel designs. Specifically, the global branch leverages singular value decomposition (SVD) for canonical alignment and establishes rotation-robust representations. Unlike methods relying on simple MLP for global encoding, we introduce a self-attention unit to enhance topological modeling and integrate a novel Graph Feature Extraction module for hierarchical feature encoding. In the global branch, we design a rotation-invariant geometric descriptor based on composite triangular structures, capturing 22-dimensional features (i.e., Euclidean distances, surface geometric angles, normal-derived angles) to overcome limitations of prior descriptors. An adaptive dual-attention fusion mechanism dynamically integrates features from both branches at each network stage, effectively preventing feature dilution. RiDBNet achieves accuracy of 93.3% and 80.5% on ModelNet40 and ScanObjectNN for the classification task, and 82.2% mIOU on ShapeNet for the segmentation task, and remains highly competitive against state-of-the-art methods.</p>