Feature Fusion Based on Multi-Scale Cascaded Attention for Radar Signal Modulation Recognition

Modulation recognition of radar signals is a core challenge in electronic countermeasures under complex electromagnetic environments. Deep learning demonstrates significant promise for signal analysis. Yet real-world applications often involve a broad Signal-to-Noise Ratio (SNR) range, spanning from low to high levels. Such scenarios present dual challenges: key feature loss and inflexible feature capture. To tackle these, we propose a progressive solution. It integrates a channel-level fusion mechanism for multiple time-frequency images (TFIs) and multi-scale dynamic cascaded attention (MDCA) module. First, the channel-level fusion mechanism reconstructs TFIs into combined feature maps for enhanced representation. Meanwhile, the MDCA uses variable-sized window groups across layers to capture fine-grained local and global features. Experiments on eight typical radar signals (SNR range from -21 dB to 3 dB) show the proposed method’s average accuracy outperforms traditional models. This verifies the engineering practicality of our approach in radar signal modulation recognition scenarios.