Convolutional Neural Networks （CNNs） have emerged as the dominant algorithms for object recognition in optical images. However， due to the complex imaging mechanism of Synthetic Aperture Radar （SAR）， labeled data is scarce and annotation costs are high， making it difficult to meet the demand of CNNs for large-scale， high-quality training data. Therefore， leveraging simulated or optical images for model training， and employing Unsupervised Domain Adaptation （UDA） techniques to bridge the domain gap between simulated or optical images and real SAR images， has become a viable solution. Nevertheless， existing UDA methods typically assume that all features from the source domain are transferable， overlooking the fact that domain-specific features may induce negative transfer. Meanwhile， noise inherent in pseudo-labels can also degrade the efficacy of domain alignment. To address these challenges， this paper proposes a domain adaptation framework that integrates feature disentanglement with label noise suppression. The framework employs a feature disentanglement module to decompose sample representations into transferable domain-invariant features and domain-specific features. By aligning only the domain-invariant features， the risk of negative transfer is effectively mitigated. Furthermore， a Weighted Generalized Cross Entropy （WGCE） loss function is designed to suppress noise interference arising during the iterative pseudo-labeling process. Experimental results on cross-domain SAR target recognition tasks demonstrate that the proposed method significantly enhances both recognition accuracy and the robustness of domain adaptation.