Automatic Modulation Classification Based on CNN and Multiple Kernel Maximum Mean Discrepancy

Abstract

Automatic modulation classification plays a significant role in numerous military and civilian applications. Deep learning methods have attracted increasing attention and achieved remarkable success in recent years. However, few methods can generalize well across changes in varying channel conditions and signal parameters. In this paper, based on an analysis of the challenging domain shift problem, we proposed a method that can simultaneously achieve good classification accuracy on well-annotated source data and unlabeled signals with varying symbol rates and sampling frequencies. Firstly, a convolutional neural network is utilized for feature extraction. Then, a multiple kernel maximum mean discrepancy layer is utilized to bridge the labeled source domain and unlabeled target domain. In addition, a real-world signal dataset consisting of eight digital modulation schemes is constructed to verify the effectiveness of the proposed method. Experimental results demonstrate that it outperforms state-of-the-art methods, achieving higher accuracy on both source and target datasets.