A Novel Lightweight Grouped Gated Recurrent Unit for Automatic Modulation Classification

With the aim of enhancing the performance of Automatic Modulation Classification (AMC), modern researchers tend to adopt architectures based on Recurrent Neural Networks (RNNs) to better extract temporal features from signals. However, the widely adopted vanilla RNN structures incur high parameter and compute overheads, making them challenging to deploy in resource-limited scenarios. In this letter, a plug-and-play architecture based on Gated Recurrent Unit (GRU) is proposed, named Lightweight Grouped Gated Recurrent Unit (LGGRU), which proves highly effective for AMC. We first analyze the detrimental impact of the reset gate in vanilla GRU on feature extraction for modulation signals and suggest its removal. Additionally, a grouped update algorithm for hidden state is innovatively proposed to further reduce redundant parameters. Furthermore, LeakyReLU is employed instead of the hyperbolic tangent (tanh) activation function, making it more adaptable to resource-limited devices. Experimental results confirm the improvement in AMC accuracy with the proposed method, along with a significant reduction in parameter and compute overheads.