In communication systems, the modulation classification of high order quadrature amplitude modulation signals is a difficult problem. An improved Transformer deep learning modulation classification method is proposed in this paper. The network parallelizes two Transformer encoders. In the additive Gaussian white noise channel, the automatic modulation classification effect of 10 modulation formats ranging from 4 QAM to 4 096 QAM with SNR ranging from -10 dB to 30 dB was analyzed. First, the quadrature and in-phase components of the QAM signal were extracted and preprocessed. Then the preprocessed in-phase component and quadrature component pass through two Transformer encoders to extract component features. Finally, the two extracted component features were combined to judge the modulation format of the QAM signal. The experimental results prove that the network can accurately identify 10 kinds of QAM modulation formats when there is no influence of carrier frequency offset and the signal-to-noise ratio is greater than 20 dB. When the carrier frequency offset is 500 Hz and the signal-to-noise ratio is greater than 26 dB, the classification accuracy of the 10 kinds of QAM modulation formats is higher than 98.6%.