A Data Preprocessing Method for Automatic Modulation Classification Based on CNN

Abstract

As a backbone of deep learning models, convolutional neural networks (CNNs) are widely used in the field of automatic modulation classification. Nevertheless, we speculate that the forms of signal samples make them inefficient for direct use as a CNN input. In this letter, a novel data preprocessing method is proposed to markedly improve CNN-based automatic modulation classification. The benchmark dataset used in this research is the well-known RadioML2016.10a dataset. The experimental results show that using the proposed method gains approximately 10% accuracy improvement in a simple CNN. Furthermore, according to the form of the preprocessed data, we designed a CNN with residual blocks to reach a maximum accuracy of 93.7% when the signal-to-noise ratio is 14 dB, which outperforms state-of-the-art automatic modulation classifiers.