Recently, specific emitter identification (SEI) has been proposed to improve communication security by identifying radio frequency fingerprint (RFF). However, complex practical environments pose problems for SEI: real signals, filled with varying noises and Doppler shifts, make SEI difficult; decreasing dynamic interference requires complex systems and real-time tuning; and accurately labeling bad signals for training is hard. To address these issues, we design an RFF template consisting of an RFF extractor complex-valued discrete wavelet transformation (CVDWT) and a classifier with flexible parameters, adaptively adjusting parameters while the disturbance changes; a complex-value particle swarm optimization (CPSO), adjusting the RFF template to build a lighter and more robust SEI system; and a special complex-valued GAN (CVGAN), reducing noises and Doppler shifts for complex-valued signals and avoiding dependence on labeling. To evaluate our system, we build a training set with only signal to noise ratio (SNR) 10dB and 0MHz shifts and a mixed test set with SNR $\in$ [-20,10] dB and [0,60] MHz shifts. For the test set with changing disturbance, our system, with fewer parameters, achieves a test accuracy of 88.5% under a 37.5MHz shift and 79.5% under -10dB, while other systems almost lose the ability to recognize the signals.