Enhanced Modulation Recognition Through Deep Transfer Learning in Hybrid Graph Convolutional Networks

Abstract

Nowadays, wireless communication plays a pivotal role in our daily lives, encompassing technologies such as wireless fidelity (Wi-Fi) and the internet of things (IoT). The backbone of the wireless communication is modulation, which involves various techniques with its own unique characteristics. As modulation techniques evolve in intricacy and diversity, the need for modulation recognition becomes apparent. Traditional modulation recognition relies on human intervention to classify modulation types in received signals, a time-consuming and laborious process prone to human error and inefficiency. Consequently, automatic modulation recognition (AMR) is introduced to autonomously classify modulation types without human interventions. In the current era, artificial intelligence (AI), specifically deep learning (DL) has gained prominence, providing numerous advantages across various domains, including AMR. While many DL-based AMR models have been developed, their efficacy reduces at low signal-to-noise ratio (SNR). Consequently, we propose a hybrid DL model for AMR, named the in-phase and quadrature - temporal graph convolutional network (IQ-TGCN) to enhance the recognition performance at low SNR. Integrating graph convolutional network (GCN) and long short-term memory (LSTM) architectures, the IQ-TGCN takes a node feature matrix as input, derived from the magnitude differences between each node. In comparative assessments against other DL models, our model has consistently exhibited superior performance. To enhance its capabilities further, we integrated deep transfer learning, leading to a remarkable 30% improvement in classification accuracy. Notably, at a SNR of 10 dB, IQ-TGCN reached its pinnacle, attaining an impressive accuracy of 99%, all the while significantly reducing training time by nearly threefold.