A lightweight multi-feature fusion structure for automatic modulation classification

Abstract

Automatic modulation classification (AMC) is a useful technology for Internet of Things devices to automatically distinguish the modulation scheme of the received signal. Recently, deep learning methods have been a popular solution for AMC, but these methods generate unsatisfactorily high model complexity while achieving acceptable classification accuracy. Therefore, developing an AMC method that can balance classification accuracy and model complexity is a challenging task, especially when targeting resource-constrained devices. In this paper, we propose a lightweight multi-feature fusion structure (lightMFFS) for AMC. This structure consists of three parts, namely data processing, feature extraction, and classification. Specifically, in the data processing part, we adopt three types of signal data, including IQ data, AP data, and IQ-transformed data. In the feature extraction part, we design three lightweight extractors to extract different spatial features of the signal from different data. In the classification part, an attention fuser is used to integrate all the features and enhance the important channel features to provide comprehensive and effective signal knowledge for classification. Extensive comparative experiments with current popular models show that our method achieves higher classification accuracy with fewer training parameters and has a more stable and superior learning capability.