Aiming at the problem that the identification accuracy of current small sample electromagnetic signal identification algorithm is low under different SNR, a Fuzzy Entropy Empirical Mode Decomposition (FEEMD) algorithm is proposed to extract electromagnetic signal feature. Extract the data with obvious characterization to expand short-time Fourier transform (STFT), and then select Transformer model to classify and identify each standard signal. The algorithm uses the recognition accuracy of 8 kinds of communication signals under the five SNR of -10dB, -5dB, 0dB, 5dB and 10dB respectively to determine the optimal hyperparameters of the network. Simulation results show that under the five SNR, the recognition rate of the four modulated signals (2FSK, AM, ASK and SSB) is more than 90%, and the accuracy rate of QAM16, QPSK and OFDM is increased from 30%-40% to more than 70%, which shows the effectiveness and practicability of the algorithm.