Modulation Classification: Diverse Techniques and Constellation Model Training at Varying Resolution

In both wired and wireless scenarios, modulation techniques play a crucial role in enhancing QoS, ensuring robustness against fading channels, reducing BER, and supporting various applications. In this work, various models, including CNN, LSTM, and ResNet, were trained on the RadioML dataset with IQ time-series, amplitude-phase, and constellation diagrams as input features. Our results indicate that time-series features with the hybrid CNN-LSTM model outperformed traditional CNN models, particularly under low SNR conditions, while amplitude-phase time-series enhanced classification performance at high SNRs. However, performance across datasets with varying channel conditions showed limited adaptability. Notably, combining time-series and constellation inputs demonstrated potential for enhanced generalization, paving the way for robust real-world modulation recognition.