Modulation recognition based on adaptive denoising residual neural network

The existing deep learning methods cannot accurately identify the modulation type of wireless signals in the circumstance of low signal-to-noise ratio (SNR). Therefore, the model based on adaptive denoising residual neural network is designed in this paper. At the input end, the pre-processing of the communication signal is carried out to extract the in-phase, orthogonal, amplitude, phase and frequency components of the complex signal as the recognition features. Then adaptive denoising module combining the soft threshold function and the improved channel attention mechanism removes the noise mixed in the signal. After that, the Inception-ResNet which is composed of Inception and residual structure is used to mine deep features. Finally the modulation recognition is realized by the full connection layer. The simulation results show that under the background of Gaussian white noise, its average recognition accuracy for common signals reaches about 94% and also performs well in the condition of Rician, Rayleigh and even unknown channel. Besides, the proposed model is better than visual geometry group-19 under low SNR, and has certain robustness to frequency offset and phase offset.