Dual-channel hybrid neural network for modulation recognition

Abstract

In the blind recognition of wireless communication modulation based on deep learning, how to further improve the recognition accuracy has become a key research issue. In this work, we propose a dual-channel hybrid model termed as CLDR (convolutional long short-term deep neural and residual network). The CLDR consists of the convolutional long short-term deep neural network (CLDNN) and the residual network (ResNet), where CLDNN is to reduce the variations in the spectrum and time, ResNet is to avoid gradient vanishing or exploding. In addition, we design an exponential curve decay adaptive cyclical learning rate method to decrease the training time cost of the neural network model. This method eliminates the need to experimentally search the optimal learning rate as with the fixed learning rate policy. It also avoids the slow convergence of the model due to the excessive attenuation amplitude as with the triangular learning rate policy. We test the feasibility of the CLDR model and discuss the influence of the exponential decay cyclical learning rate on the training of CLDR model based on the RadioML2016.10b public dataset. Simulation results show that the CLDR model yields a recognition accuracy of 93.1% at high SNRs, which effectively reduces the influence of external environment such as noise and fading on recognition accuracy. The training time cost of CLDR using exponential cyclical learning rate is reduced by 14.6% and 32.1% compared with triangular and fixed methods. Therefore, the exponential cyclical learning rate policy effectively reduces the training time cost of the model in the same classification accuracy.