Multilevel Adaptive Wavelet Decomposition Network-Based Automatic Modulation Recognition: Exploiting Time-Frequency Multiscale Correlations

Deep Learning-enabled Automatic Modulation Recognition (DL-AMR) is a pivotal technique for designing intelligent communication receivers. However, conventional DL-AMR approaches often fail to fully adapt neural networks to wireless communication signals’ inherent properties, resulting in reduced recognition accuracy or increased network complexity. To address these challenges, this paper proposes a Multilevel Adaptive Wavelet Decomposition Network (MAWDN) that leverages time-frequency multiscale correlations. The architecture of MAWDN comprises three key modules: Multi-channel feature extraction, adaptive wavelet decomposition, and residual classification flow. The network begins by processing I/Q signals through a multi-channel convolutional neural network to extract temporal features and inter-channel correlations. These features are then processed through an adaptive wavelet decomposition module to enhance frequency characteristic analysis. The final stage employs a residual classification flow to effectively integrate information across various wavelet scales. We evaluated our model on the RML2018.01a and HisarMod2019.1 datasets. MAWDN performed exceptionally well on both datasets, demonstrating strong adaptability across different datasets. Compared to the baseline models, MAWDN significantly improved classification accuracy without substantially increasing model complexity. Ablation studies confirm the positive impact of each module on the overall performance. Additionally, this paper includes visualizations of the feature extraction process, providing an intuitive understanding of the model’s operational dynamics.