A transformer-based framework with complex-valued convolution and enhanced Bi-LSTM for automatic modulation recognition

Abstract

Automatic Modulation Classification (AMC) is critical for modern wireless communication systems. Despite significant progress in deep learning-based AMC, existing methods still struggle to jointly capture local features, model temporal dependencies, and extract global representations. We propose CBADNN, an end-to-end architecture that combines Transformer self-attention for global context modeling, complex-valued convolutions for local spatial feature extraction, and bidirectional stacked LSTMs (Bi-sLSTMs) for temporal dependency learning. CBADNN achieves state-of-the-art performance, with overall accuracies of 64.02% and 65.50% on the widely used RadioML 2016.10a and RadioML 2016.10b datasets, respectively. On RadioML 2016.10a, it outperforms the best baseline by 0.81% under high SNR (4 dB–18 dB) and 0.33% under medium SNR (−8 dB–2 dB), consistently demonstrating superiority across diverse SNR conditions.Furthermore, a detailed evaluation of classification accuracy across various modulation types under different SNR conditions is conducted, and ablation experiments on the RadioML2016.10a dataset are performed to rigorously validate the effectiveness of each module in the proposed network.