Hyper-relational Knowledge Graph (HKG) link prediction is a critical research area with substantial academic and practical significance. HKG consists of hyper-relational facts, comprising a primary triple augmented by multiple attribute-value qualifiers, enabling rich factual representation. However, noise is inevitably introduced during knowledge graph construction. Existing methods often fail to simultaneously ensure link prediction accuracy and robustness, thereby limiting their effectiveness in downstream applications. To address this issue, we propose RMNS, a novel approach based on a heterogeneous graph encoder. RMNS employs a multi-level negative sampling strategy to perform forward diffusion and backward denoising on all related elements, with an emphasis on low-confidence components, thereby improving model robustness and effectively suppressing noise. Additionally, RMNS incorporates an edge-biased attention mechanism to differentially emphasize heterogeneous element embeddings, enabling more precise capture of associations within hyper-relational structures. Experimental results on JF17K and Wikipeople benchmark datasets demonstrate that RMNS improves the Hits@1 index for entity and relation link prediction by an average of 2.3% and 0.4%, respectively. It is verified that this method has significant advantages in improving the accuracy of link prediction in the case of noise interference.