HFDNN: A Hybrid Fusion Deep Neural Network for Robust Automatic Modulation Classification in Adverse Wireless Environments

Automatic Modulation Classification (AMC) plays a vital role in modern wireless communication systems, particularly in domains such as defense, the Internet of Things (IoT), and cognitive radio, where resilience to interference is essential. Traditional AMC methods often struggle under real-world impairments, including Additive White Gaussian Noise (AWGN), Rayleigh fading, jamming, and hardware non-idealities. To address these limitations, we present a Hybrid Fusion Deep Neural Network (HFDNN) that combines the strengths of VGG, LSTM-CNN, GRU-CNN, and Convolutional Long Short- Term Memory Deep Neural Networks (CLDNN) architectures. This fusion enables the model to effectively capture both spatial and temporal features across a broad range of Signal-to-Noise Ratios (SNRs). The model is trained on a diverse set of modulation schemes, Amplitude Shift Keying (ASK), Phase Shift Keying (PSK), Amplitude Modulation (AM), Frequency Shift Keying (FSK), Amplitude and Phase Shift Keying (APSK), and Quadrature Amplitude Modulation (QAM) and evaluated under various channel impairments such as Carrier Frequency Offset (CFO), phase noise, and fading. Experimental results demonstrate that HFDNN consistently outperforms traditional deep learning models. It achieves 66.72% accuracy at −20 dB, exceeds 97% beyond −2 dB, and reaches 89.13% accuracy with transfer learning. Under channel interference, classification accuracy improves by 30–40% at low SNRs and 10–15% at high SNRs, significantly narrowing the gap with interference-free conditions. In addition to superior accuracy, HFDNN provides an effective balance between prediction time (8.56 ms) and F1-score (0.89), making it suitable for real-world deployment.