Lightweight Deep Learning Model for Automatic Modulation Classification in Cognitive Radio Networks

Abstract

Automatic modulation classification (AMC) used in cognitive radio networks is an important class of methods apt to utilize spectrum resources efficiently. However, conventional likelihood-based approaches have high computational complexity. Thus, this paper proposes a novel convolutional neural network architecture for AMC. A bottleneck and asymmetric convolution structure are employed in the proposed model, which can reduce the computational complexity. The skip connection technique is used to solve the vanishing gradient problem and improve the classification accuracy. The dataset DeepSig:RadioML, which is composed of 24 modulation classes, is used for the performance analysis. Simulation results show that the classification accuracy performance of the proposed model is outstanding in the signal-to-noise ratio (SNR) range from -4 dB to 20 dB compared with MCNet that is the best model in the conventional models, where the proposed model achieves 5.52% and 5.92% improvement regarding classification accuracy at the SNRs of 0 dB and 10 dB, respectively. In terms of the computational complexity, the proposed model not only saves the trainable parameters by more than 67% but also reduces the prediction time for a signal by more than 54.4% compared with those of MCNet.