Efficient Hybrid Neural Network for Automatic Modulation Recognition

Automatic Modulation Classification (AMC) is a fundamental task to blindly identify modulation schemes within Radio Frequency (RF) signals. This process holds paramount importance within the realm of Cognitive Radio (CR) applications, enabling critical tasks such as interference detection and link adaptation. Deep learning (DL) has shown remarkable effectiveness in addressing the AMC challenge; however, most DL-based AMC schemes have high processing and storage requirements, rendering them unsuitable for resource-constrained devices. To tackle this issue, this paper introduces a lightweight neural network (NN) constructed by fusing Gated Recurrent Units (GRUs) and multiple convolutional blocks. These convolutional blocks are meticulously designed using asymmetric kernels to reduce computational complexity, and Squeeze and Excitation (SE) blocks to enhance channel interdependencies with minimal computational cost. Additionally, the convolutional blocks incorporate skip connections to improve classification accuracy and mitigate the vanishing gradient problem. Experimental results on the RadioML 2016.10A dataset demonstrate that our model outperforms benchmark DL-based AMC models in terms of accuracy while utilizing the fewest trainable parameters and exhibiting moderate inference time.