<![CDATA[Automatic Modulation Classification (AMC) is a pivotal technology for efficient spectrum management in future cognitive radio networks. While Deep Learning has advanced the field, standard Convolutional Neural Networks (CNN) often struggle to capture long-term temporal dependencies in signals affected by fading. This study proposes an Optimized Hybrid CLDNN architecture that integrates a "Wide-Kernel" CNN (k=7) for enhanced spatial feature extraction and a "High-Capacity" LSTM (100 units) for robust temporal modeling. Experimental validation using the RadioML 2016.10a dataset demonstrates that the proposed optimizations yield significant performance gains. Specifically, the model achieves a classification accuracy of 84.5% at 0 dB SNR, outperforming standard baselines in the critical transition regime. Furthermore, it reaches a peak accuracy of 92.4% at high SNR (+18 dB). A notable finding is the reduction of inter-class confusion between 16-QAM and 64-QAM, where the misclassification rate is suppressed to approximately 15%, confirming the architecture's effectiveness in resolving hierarchical modulation ambiguities in dynamic wireless environments.]]>
Copyrights © 2026
