Current multimodal sentiment analysis （MSA） models predominantly employ cross-modal attention mechanisms to process feature information from different modalities. However， these approaches often overlook the inherent similarities and dissimilarities among modal features， which can easily lead to the generation of redundancy from modal similarities and an increase in noise during cross-modal interaction， thereby degrading model performance. To address these issues， this paper proposes a novel multimodal sentiment analysis model based on feature decoupling and cross-modal deep interaction enhancement （FD-CMDIE）. Firstly， for feature extraction， NeoBERT is utilized to extract high-quality textual features， while stacked long short-term memory （LSTM） networks are employed for visual and acoustic features. Subsequently， common and private encoders are used to decouple the features of the three modalities into similar and dissimilar features. Contrastive learning is then applied， using the textual similar features as anchors， to pull similar features from different modalities closer in the feature space while pushing dissimilar features further apart. Finally， a cross-modal interaction enhancement network is designed for deep interaction and fusion of the decoupled features， and a gated attention pooling module is utilized to filter out noise generated during the interaction. Experiments conducted on two benchmark datasets demonstrate that our proposed method surpasses several state-of-the-art approaches across most metrics， validating its effectiveness.