Communication Modulation Classification based on multi modality feature deep learning

Automatic Modulation Classification (AMC) is a crucial step before signal analysis in non-cooperative scenarios and an important component of cognitive radio technology. Recently, AMC methods based on Deep Learning (DL) have received increasing attention due to their powerful feature extraction capabilities. However, most existing DL-based methods only consider a single feature as input, without leveraging different attributes of the signal, which limits classification performance. To achieve better classification performance, we propose a cross-domain multimodal scheme for AMC. This scheme sends features from different modalities into a single DL model and controls their effects by multiplying weights. The proposed method divides different modal domains into three groups, including raw data, time domain, and frequency domain. Eight modal features are selected, and an analysis of different modal features is conducted. Finally, we investigate the complementarity between different modal features through extensive simulation experiments, which helps guide how to select feature modal transformations based on the modulation set, thereby effectively extracting knowledge from the received signal.