Abstract: The identification performance with convolutional neural network (CNN) is limited with the types of signal modulation identification. For instance, the identification accuracy is just only 80% when 24 kinds of modulation waveforms presented at SNR=12 dB. If a better recognition performance wants to be obtained, more complicated network must be required. It directly enlarges the requirement of the data set size and the cost of hardware calculation resources is also increased. Therefore, a compact residual neural network for radio signal modulation identification is designed in the paper, which can be used to extract the characteristics of signal modulation. The end-to-end identification is accomplished from the baseband in-phase and quadrature components. By using transfer learning, the number of samples in the network retraining stage is reduced dramatically and the adaptive ability of the proposed network is enhanced. The test results show that the identification performance with the proposed neural network approaches 95% when the SNR is 12 dB even though the wireless channel impulse response verified. Several comparative experiments illustrate the advantages of the proposed neural network.

Key words: deep modulation identification (DMI), residual neural network (ResNet), transfer learning (TL), data-driven, convolutional neural network (CNN)