Modulation recognition is a critical task in wireless communications. Although deep learning methods have achieved remarkable progress in this field, they still face the challenge of insufficient generalization ability in complex non-cooperative environments—particularly when confronted with varying channel conditions, which can obscure the subtle discriminative features between structurally similar modulation schemes (e. g. 16QAM and 64QAM) and thus degrade recognition performance. To address this unique challenge in the field of modulation recognition, an unsupervised adversarial domain adaptation method named Feature Alignment and Discrimination Domain Adaptation ( FADDA) is proposed. The core of FADDA is the introduction of a contrastive learning-based feature alignment loss on the basis of adversarial training. Adversarial training is responsible for learning domain-invariant features to adapt to channel variations, while the feature alignment. loss fundamentally enhances the model' s ability to distinguish between easily confused modulation types by explicitly reinforcing the compactness of intra-class features and the separability of inter-class features. Experimental results show that without target-domain labels, this method can significantly improve the model's cross-channel modulation recognition performance and demonstrate strong generalization ability.