Multi-Feature Fusion for Enhanced Feature Representation in Automatic Modulation Recognition

Modulation recognition plays a crucial role in the efficient management of spectrum resources. However, traditional methods have long posed challenges for researchers due to their excessive reliance on manual effort. With the advancement of deep learning, automatic modulation recognition has emerged as a promising solution. Nonetheless, most existing deep learning-based modulation recognition studies consider only single-domain feature information of the signal, such as time-domain or frequency-domain features. To further enhance recognition performance, this paper proposes a novel multi-feature input fusion network. By utilizing different representations and processing methods of the signal, the proposed approach designs distinct feature extraction networks tailored to specific processed signals, leveraging the characteristics of convolution kernels with varying sizes and receptive fields. Additionally, the attention mechanism module is improved to enhance network performance while minimizing the increase in parameter count. Experimental results on the publicly available RML2016.10a dataset demonstrate that the proposed model achieves highly efficient recognition above 2 dB, with accuracy approaching 100%. Testing with real-world signals collected under realistic channel conditions using USRP devices further confirms the model's superior performance, achieving an effective recognition rate of 99.27% when the two USRP devices are in close proximity. These results validate the model's high efficiency and robustness in both simulated and real-world environments.