Automatic Modulation Recognition for Radar Signal Based on Multi-Level and Multi-Scale Feature

With the increasing adoption of deep learning in signal processing, the performance and efficiency of automatic modulation recognition (AMR) for radar intra-pulse signal have been significantly enhanced. However, existing methods typically focus on feature extraction at a single level, neglecting the potential of multi-level feature learning. Consequently, they are often inadequate in addressing the complexity and diversity of radar signals. This paper proposes a Multi-level and Multi-scale Feature learning approach for AMR (MMFAR) that combines masked autoencoder and contrastive learning to improve the accuracy and robustness of radar signal modulation recognition. During the pre-training phase, the model generates positive sample pairs by applying random masking twice. Subsequently, these samples are processed through a dual-path: one path is dedicated to signal reconstruction, while the other focuses on feature comparison. The reconstruction path uses mean squared error loss for optimization at the sampling point level, ensuring the quality of signal recovery, while the feature extraction path applies contrastive loss for learning discriminative features at the feature level. Finally, the model is fine-tuned through cross-entropy loss to further enhance classification performance at the sample level. Furthermore, the encoder employs convolutional networks to extract multi-scale features and integrates a multi-head attention mechanism to capture long-range dependencies in complex signals. Experimental results demonstrate that the proposed approach significantly outperforms existing methods in modulation recognition tasks under low signal-to-noise ratio (SNR) and complex environmental conditions.