Abstract: Radio frequency fingerprint （RFF） recognition is a key technology for target identification， enabling precise recognition of individual unmanned aerial vehicles （UAV） based solely on physical-layer signal characteristics without the need for decoding or decryption. Subtle variations in the manufacturing process of UAV radio frequency front ends lead to different devices exhibiting unique hardware nonlinearities， frequency responses， and phase noise characteristics， thereby forming intrinsic RFF signatures that can be used for individual differentiation. However， conventional RFF recognition methods typically rely on single-domain feature representations， leading to limited anti-interference capability and inadequate modeling of dynamic signal characteristics. To address these challenges， this paper proposes an efficient multi-domain feature fusion method that deeply integrates in-phase/quadrature， bispectral， and short-time Fourier transform spectrogram features to enhance RFF representation. The proposed fusion architecture effectively improves recognition accuracy while enhancing the model’s generalization capability. To overcome the computational complexity and real-time bottlenecks introduced by multi-domain fusion， an efficient fusion framework based on Mamba and Vision-Mamba architectures is further developed. This framework leverages the linear-time computational complexity of Mamba for long-sequence modeling and incorporates hardware-aware optimization strategies， thereby significantly reducing computational overhead while maintaining high recognition accuracy. Furthermore， to balance feature redundancy and complementarity during fusion， an adaptive cross-attention mechanism is introduced to dynamically model the interdependencies among multi-source features. This mechanism enables adaptive weighting of feature contributions， effectively mitigating the feature degradation caused by environmental interference. Experimental results demonstrate that the proposed method achieves a recognition accuracy of 97.76% while maintaining low computational cost and fast inference speed， confirming its effectiveness and superiority in UAV RFF identification.