<p>Multimodal Sentiment Analysis (MSA) aims to capture complex emotions by integrating diverse modalities. However, existing methods often struggle to distinguish cross-modal consensus from conflicts, while facing challenges from redundant noise and feature manifold distortion. To address these issues, we propose FDC-MSA, a framework for noise suppression and conflict decoupling in inconsistent MSA. The model enhances sentiment modeling through three integrated components. First, a Variational Information Bottleneck (VIB) strategy is employed to compress and reconstruct auxiliary semantics. By suppressing task-irrelevant redundancy while maximizing sentiment sufficiency, it achieves semantic purification. Second, the Geometric Orthogonal Decoupling Routing (GODR) mechanism decomposes visual representations into consensus and specificity components, utilizing dynamic routing to adaptively model cross-modal conflicts. Furthermore, a Prototype-Guided Topology Rectification (PGTR) module stabilizes high-dimensional representation spaces by guiding feature distributions toward category structures. Experimental results on MVSA-Single, MVSA-Multiple, and HFM datasets demonstrate that FDC-MSA achieves state-of-the-art performance, notably reaching 78.96% accuracy on MVSA-Single. The framework exhibits superior robustness in handling text-image sentiment inconsistencies, validating the effectiveness of our denoising and decoupling strategies in complex cross-modal scenarios.</p>