Modulation Recognition Based on Multi-modal Fusion of ResNet and Traditional Signal Processing

Modulation recognition is a key technology in wireless communications, particularly valuable in non-cooperative communication scenarios. Traditional methods face challenges such as high computational complexity and insufficient accuracy at low signal-to-noise ratios. While convolutional neural networks enable automatic feature learning, they tend to suffer from gradient vanishing issues as network depth increases. Residual neural networks address this limitation through shortcut connections while enhancing hierarchical feature extraction capabilities. This paper innovatively proposes a multi-modal fusion approach combining traditional signal processing with ResNet-18 deep learning. By leveraging the advantages of both methods under resource-constrained conditions, we significantly enhance constellation diagram features through signal preprocessing, achieving model lightweight optimization without sacrificing accuracy, thus balancing real-time performance and precision. Our research establishes a wireless communication experimental system based on the ADALM-PLUTO SDR platform for processing and recognizing modulated signals in physical channel environments. Experimental results demonstrate that when using the ResNet-18 neural network model for modulation recognition, it exhibits significant speed advantages under equivalent channel conditions. Compared to the ResNet-50 network, ResNet-18’s inference time is only 49.02% of the former, demonstrating its computational efficiency advantage over deeper network architectures. Meanwhile, it achieves recognition accuracy exceeding 98%, providing valuable reference for research in this field.