A Novel Approach for Multicarrier Modulation Signals Using GTCC Spectrogram and Improved Transfer Learning Models

Abstract

High-speed and low-latency communication are critical for next-generation wireless networks, demanding accurate and computationally efficient signal demodulation. However, conventional demodulation techniques suffer from high complexity and poor robustness under low signal-to-noise ratio (SNR) conditions, limiting real-time applicability. This study presents a novel AMC framework for 5G systems using simulated multicarrier modulation (MCM) signals from the Vienna 5G Link Level Simulator with Jake's Doppler model and varying SNRs via added Gaussian noise. The front end of the proposed method uses an innovative feature extraction approach using gammatone cepstral coefficient (GTCC) spectrograms that capture fine-grained spectral–temporal patterns of MCM signals, enhancing feature discrimination. At the back end of the proposed scheme, three enhanced deep learning models, Improved-InceptionV3, Improved-Xception, and Improved-ResNet152V2, have been applied, individually, integrating fine-tuned convolutional layers, dense blocks, dropout regularization, and adaptive learning rates for superior performance. Experiments on the 5G waveform dataset demonstrate state-of-the-art classification accuracy, with Improved-ResNet152V2 leveraging architectural pruning and quantization outperforming Improved-InceptionV3 and Improved-Xception, achieving 96% at −16 dB SNR and 100% at 16 dB SNR. The synergy of GTCC spectrograms and enhanced architectures enables robust accuracy–efficiency trade-offs, outperforming existing methods in low SNR and noisy conditions, while Grad-CAM visualizations enhance interpretability and reliability for AMC tasks. This framework establishes a significant advancement in AMC research.