Automatic modulation recognition using deep CVCNN-LSTM architecture

Abstract

Automatic modulation recognition (AMR) stands as a pivotal operation within industrial cognitive radio systems. State-of-the-art real-valued convolutional neural networks (CNNs) have innovated modulation recognition but view complex signal components as separate real inputs, impairing signal phase integrity and model interpretability. This paper presents an innovative AMR method known as the CVCNN-LSTM. Our study aims to leverage CNNs in conjunction with a long short-term memory network (LSTM) to harness the strengths of both networks while mitigating their weaknesses. This approach enhances the learning of original I/Q signal features, thus improving recognition performance. The proposed network is composed of multiple LFLBs and an LSTM layer. Each LFLBs is primarily composed of a convolutional layer and an average pooling layer, which play a key role in enabling local correlation learning and hierarchical correlation extraction. Recognizing that I/Q data inherently possesses a complex-valued structure, we advocate using end-to-end complex-valued CNN and complex-valued LSTM instead of dual-channel real-valued networks for modulation recognition. We conduct experiments to evaluate complex-valued networks and demonstrate that our method surpasses current state-of-the-art approaches in the AMR field.

Keywords

• Modulation recognition
• CNN
• LSTM
• Complex-valued
• I/Q signals