In the context of unsupervised domain adaptation (UDA) for point cloud classification, deep classifiers training on data from source domain (e.g. clean synthetic point clouds) cannot perform well on those from target domain (e.g. noisy real-world ones), which can be caused by a significant domain discrepancy of point representations. For closing domain gap, recent algorithms adopt the popular self-training strategies (e.g. self-paced self-training) but suffer from lack of imposing structural constraints into representation learning. To tackle this issue, we propose a novel dual-augmentation relational learning scheme (i.e. introducing consistency regularization on augmented samples in both observation and feature space) to incorporate low-dimensional manifolds to encourage domain-invariant representations. Moreover, we design a novel filtering mechanism that adaptively adjusts thresholds for each semantic category based on confidence distributions and validates neighborhood consistency to further mitigate feature ambiguities. Comprehensive experiments on the widely-used PointDA-10 dataset demonstrate that our method achieves the state-of-the-art performance.