Automatic modulation classification using recurrent neural networks

Abstract

Automatic modulation classification (AMC) is one of the essential technologies, and also a hard nut to crack in the field of cognitive radio (CR) and non-cooperative communication systems. In this work, we propose a novel AMC method based on the promising recurrent neural network (RNN), which is shown to have the capability to sufficiently exploit the temporal sequence characteristic of received communication signals. This method resorts to raw signals directly with limited data length, and avoids extracting signal features manually. The proposed method is compared with a convolutional neural network (CNN) based method and the result indicates the superiority of the proposed one, especially when signal-to-noise ratio (SNR) is above -4dB. Furthermore, a comparative study is presented to evaluate the availability of the other different RNN structures. And a more efficient structure is recommended based on two-layer gated recurrent unit (GRU) network. Additional numerical results demonstrate that the proposed structure achieves an improved performance from 80% to 91% in terms of classification accuracy.