A Spatial-Energy-Frequency Aware Framework for Joint Recognition and Restoration of Multipath Radar Signals

Abstract

The analysis of noncooperative radar signals is of great importance in both military and civilian applications. However, multipath fading during transmission often introduces severe signal distortions, posing great challenges to accurate signal recognition and reliable restoration. To address these issues, this article proposes a spatial-energy-frequency aware network (SEFANet) for the joint intrapulse modulation recognition and restoration of radar signals affected by multipath propagation. First, a multipath channel model is established and its impact on radar signals is analyzed. Then, the time-frequency representation of the signal is employed as input to characterize its variation patterns. To enhance feature learning under multipath distortion, two key modules are designed, namely the spatial-energy aware module (SEAM) and the selective frequency-aware module (SFAM). The SEAM adaptively adjusts convolutional kernel shapes and reweights features according to the signal’s spatial and energy distribution, thereby improving the network’s adaptability to multipath-distorted radar signals. Meanwhile, the SFAM introduces frequency-domain priors through wavelet decomposition and selectively preserves dominant structural component, further enhancing discriminative FE. By integrating these two modules, SEFANet achieves effective feature representations in a unified latent feature space for both recognition and restoration tasks. Extensive experiments demonstrate the superiority of the proposed SEFANet. Compared with state-of-the-art methods, SEFANet not only achieves higher recognition accuracy but also exhibits outstanding signal restoration capability.