Unsupervised multiplex graph representation learning (UMGRL) has attracted significant attention due to its ability to capture complex interaction information effectively without relying on labeled data. However, few studies have adequately addressed the prominent noise issue while simultaneously capturing the complementarity and consistency within the multiplex graph. To alleviate the above limitation, we propose a UMGRL method, named MD3C. It designs denoising encoders by integrating multi-perspective information from original node features, multi-hop neighborhoods, and different iterations. Subsequently, it mines intra-graph denoising complementarity through reconstruction and dimensional decorrelation constraints. Additionally, it achieves inter-graph consistent consensus by the cross-graph guided contrast and correlation maximization constraints. MD3C collaboratively counteracts noise and comprehensively extracts valuable information in the multiplex graph, thereby generating high-quality representations. Extensive thorough experiments validate the effectiveness and superiority of the proposed method across various downstream tasks.