Akademik Veri Yönetim Sistemi
IEEE ACCESS, cilt.14, ss.1-36, 2026 (SCI-Expanded, Scopus)
Automatic Modulation Classification (AMC) is a fundamental task in modern wireless communication systems, particularly in cognitive radio, software-defined radio, and next-generation communicationenvironments. In this study, a two-stage hybrid Long Short-Term Memory (LSTM)-Multi-Layer Perceptron (MLP) architecture is proposed for robust modulation classification under varying Signal-to-Noise Ratio (SNR) conditions. In the proposed framework, the received signal is first represented by Inphase/Quadrature (I/Q) and Amplitude/Phase (A/P) features to provide a more discriminative input space. The first stage performs analog/digital discrimination, while the second stage activates the corresponding analog or digital sub-classifier, thereby reducing the decision space and limiting inter-class confusion. To improve representation diversity and strengthen learning under low-SNR conditions, a combined Variational Autoencoder-Generative Adversarial Network (VAE-GAN) model is employed for data augmentation. To prevent information leakage, the RadioML2018.01a dataset is first divided into disjoint 60/40 training and test subsets, and augmentation is applied exclusively to the training subset. All compared models are trained and evaluated under the same split and the same real test subset. Experimental results show that the proposed method achieves 100% first-stage accuracy from 8 dB onward, while the highest second-stage accuracies reach 99.9% for analog modulation and 98.4% for digital modulation. The findings demonstrate that the proposed two-stage hybrid LSTM-MLP model provides stable and highly accurateAMCperformance across a wide SNR range.