This paper proposes a novel Specific Emitter Identification (SEI) method based on an improved Transformer architecture to address the challenges of low recognition accuracy and weak generalization in complex electromagnetic environments. Traditional SEI methods often struggle with dynamic signal variations and overlapping emitter characteristics. To overcome these limitations, we design a multi-head adaptive attention mechanism that dynamically adjusts the weights of temporal and spectral features in electromagnetic signals, enhancing the model's ability to capture discriminative fingerprint patterns. Furthermore, a hybrid convolutional-Transformer encoder is introduced to integrate local signal characteristics with global contextual dependencies.