DGrA: Lightweight Modulation Recognition Based on Hybrid Neural Networks

Automatic modulation recognition has been recognized as an effective technique for non-cooperative communication and intelligent transmission. In this paper, we propose a new lightweight method for automatic modulation recognition, aiming to extract crucial discriminative features of signals for higher recognition accuracy while reducing spatial costs. To enhance the dissimilarity between samples, this paper combines an improved attention block and convolutional operations with the recurrent neural network, focusing on key features during the training phase to efficiently differentiate signal sequences. By replacing standard convolutions with depthwise separable convolutions, the model's computational complexity is reduced while enhancing its feature extraction capability. Furthermore, the method incorporates pruning to reduce ineffective features, decreasing the model size while maintaining performance. Experimental results on RadioML2016.10a demonstrate that the proposed method outperforms other comparative methods, exhibiting both higher recognition accuracy and smaller model size. To validate real-world applicability, the algorithm was implemented on a software-defined radio platform for signal transmission and reception under practical conditions, achieving an accuracy of 87.22% in the presence of environmental noise, thus confirming its effectiveness in real-world scenarios.