To address the difficulty of effectively extracting transient features from communication signal waveforms in communication emitter identification, as well as the lack of capability to characterize fingerprint features caused by emitter device nonlinearity, this study proposes a communication emitter identification method that combines time-frequency attention with multi-scale entropy. First, starting from the transmitter’s physical chain, a generalized model incorporating filter transient distortion and power amplifier nonlinear distortion is established to achieve dual-scale characterization of hardware features using a “local dynamic-global statistical” approach. Then, during the feature extraction phase, a time-frequency attention reconstruction mechanism is used to enhance transient abrupt-change features, while multi-scale sample entropy quantifies nonlinear complexity. Finally, a gated weighted fusion strategy is employed for adaptive fusion of heterogeneous features, and experiments are conducted using both simulated datasets and the publicly available ORACLE dataset. Experimental results show that, compared with existing methods, the proposed approach achieves emitter identification accuracies of 99.62% and 96.81% at a signal-to-noise ratio of 20 dB, which validates the effectiveness and robustness of the proposed method.