Deep Hierarchical Network for Automatic Modulation Classification

Abstract

In non-cooperative communication scenarios, automatic modulation classification (AMC) is the premise of information acquisition. It has been a difficult issue for decades due to the attenuation and interference during wireless transmission. In this paper, a novel deep hierarchical network (DHN) based on convolutional neural network (CNN) is proposed for the AMC. The model is designed to combine the shallow features with high-level features. Thus, it can simultaneously have global receptive field and location information through multi-level feature extraction and does not require any transformation of the raw data. To make full use of limited data, a new method is proposed to use signal-to-noise ratio (SNR) as a weight in training instead of working as an indicator of system robustness. Furthermore, some other deep learning methods have been used to explore whether they could improve the performance of the proposed model. Several new techniques have been chosen to be applied in the DHN at last. Then, a detailed analysis of the improvement in network performance is provided. Combination of the DHN and the weighted-loss can achieve more than 93% classification accuracy which is the best performance in an open source dataset.

Keywords

• Automatic modulation classification
• deep hierarchical network
• receptive field
• signal-to-noise ratio weight