Robust automatic modulation recognition via adaptive wavelet selective fusion architecture

Automatic modulation recognition (AMR) plays a crucial role in intelligent communication systems. Deep neural networks (DNNs), with their powerful nonlinear modeling capabilities, have significantly enhanced the performance of AMR. However, attention-based AMR models are susceptible to noise interference in low signal-to-noise ratio (SNR) conditions, where the attention distribution during feature extraction tends to focus on spurious feature regions, resulting in the failure of effective feature focusing. To address this issue, we propose an adaptive wavelet selective fusion AMR architecture (AWSF-Net). The proposed architecture employs a progressive attention refinement mechanism to dynamically optimize the feature extraction process, thereby enhancing the model’s sensitivity and attention focus on subtle modulation features. Evaluations on two public benchmark datasets indicate that AWSF-Net has approximately 0.20 million parameters. It achieves average accuracies of 86.12% and 85.95% over SNRs from -8 to 4 decibels, as well as overall accuracies of 64.42% and 65.90%, respectively. In addition, adaptive wavelet correction enhances the separability of adjacent quadrature amplitude modulation (QAM) classes.