Abstract: At present, the majority of neural network-based radar modulation signal recognition algorithms predominantly rely on a single source of information, overlooking the potential benefits that arise from the synergistic use of multi-modal information features. To tackle this limitation, a novel multi-modal parallel feature fusion model has been proposed, which leverages both one-dimensional signal sequences and two-dimensional time-frequency representations. Initially, the temporal feature extraction module incorporates a two-dimensional temporal change modeling approach to capture temporal dynamics, while the frequency domain feature extraction module employs an inverse residual structure with a band linear bottleneck layer to extract spectral features. Subsequently, the integration of two distinct attention mechanisms, along with residual connections, facilitates an effective fusion of multi-modal features, enhancing their complementary nature. Empirical evaluations conducted on DeepRadar2022 and self-built datasets demonstrate that this model provides a more comprehensive feature representation achieves higher classification accuracy and exhibits noise resilience, making it a promising solution for advanced radar signal processing applications.

Key words: radar modulation signals recognition, feature fusion, attention mechanisms, temporal two- dimensional changes, inverse residual module