Abstract: Imperfections in the analog components of emitters cause distortions in the transmitted signals. These distortions serve as unique fingerprints for the purpose of specific emitter identification (SEI). Features for SEI are typically extracted using either a distortion-models based method or a machine learning based method. In this study, these two methods for feature extraction from emitters are investigated. Incorporating specialist knowledge, the distortion model, into the neural network, a cascade network mode is proposed to extract the parameters in-phase/quadrature imbalance and phase noise models of the emitter, which not only ensures the interpretability of the extracted features and enhance identification accuracy. Simulation results demonstrate that this scheme outperforms both the conventional distortion-models-based and the machine learning-based methods in terms of feature extraction performance.