Research on Automatic Recognition and Optimization of Radio Modulation Signals Based on CLDNN Neural Networks

Modulation signal recognition is a key technology in electronic information systems, widely used in defense communications and civil spectrum management. This paper studies the CLDNN (Convolutional Long Short-Term Deep Neural Networks) model for modulation signal recognition, aiming to demonstrate its accuracy superiority over standalone CNN and LSTM. To address the issues of low recognition accuracy for complex radio signals and the large parameter volume leading to slow operation and difficult deployment, two optimization directions are proposed: a more generalizable integrated CLDNN model and a smaller parameter GRU-based CLDNN model, both achieving expected improvements. Using the DeepSig Dataset: RadioML 2018.01A, the study compares model performance under low signal-to-noise ratio (SNR) conditions, showing that CLDNN achieves 82.4% accuracy at 10 dB SNR, outperforming CNN and LSTM. The optimized integrated CLDNN model improves weighted accuracy by 4%, and the GRU-optimized model reduces training time by 59.3%. This research provides an efficient scheme for real-time modulation signal recognition, with important theoretical and application value.