Efficient Squeeze-and-Excitation-Enhanced Deep Learning Method for Automatic Modulation Classification

The rapid proliferation of mobile devices and Internet of Things (IoT) gadgets has led to a critical shortage of spectral resources. Cognitive Radio (CR) emerges as a propitious technology to tackle this issue by enabling the opportunistic use of underexploited frequency bands. Automatic Modulation Classification (AMC), which serves as a technique to blindly identify modulation types of received signals, plays a pivotal role in carrying out several CR functions, including inference detection and link adaptation. Recent research has turned to Deep Learning (DL) networks to overcome the shortcomings of traditional AMC techniques. However, most existing DL approaches are impractical for resource-limited systems. To address this challenge, we propose a novel lightweight hybrid neural network for AMC that fuses Convolutional Neural Networks (CNNs) and Gated Recurrent Units (GRUs) layers, along with a customized Squeeze and Excitation (SE) block. The integration of CNNs and GRUs allows for the learning of both spatial and temporal dependencies in modulated signals, while the SE block recalibrates features by modeling interdependencies between CNN network channels. Our experimental results, using the RadioML 2016.10A dataset, clearly demonstrate the superior performance of our approach in effectively managing the tradeoff between accuracy and complexity compared to baseline methods. Specifically, our approach achieves the highest accuracy of 91.73%, surpassing all reference models while reducing the memory footprint by at least 45%. In future work, further investigation is warranted to differentiate modulations sharing temporal or frequency domain characteristics and enhance classification accuracy in high-noise environments.