Dynamic Graph Network Augmented by Contrastive Learning for Radar Target Classification

In recent years, radar echo signals based multi-aircraft recognition has gained significant attention in the research community. Despite its potential, the extraction of valuable information from weak echo signals in complex environments remains a challenging task. Based on heterogeneous radar network, graph convolution network (GCN) is a powerful method to extract semantic features and then fuse them for reliable detection. However, owing to the limited knowledge, the design of a suitable adjacent matrix is vital for GCN. In this paper, an intelligent and partly interpretable method termed Dynamic Graph Network augmented by Contrastive Learning (CL-DGN) is proposed for classifying the radar targets in a heterogeneous radar network under low signal-to-noise ratio. The Zhao-Atlas-Marks (ZAM) transform and Gated Recurrent Unit with attention mechanism method are used for semantic feature enhancement, then the contrastive learning-based semantic similarity metric learning (SSML) is elaborated to construct an adjacency matrix to flexibly support dynamic networking for each target in dynamic heterogeneous radar network. Finally, graph convolution network method utilizes the adjacency matrix generated by SSML to extract and fuse the target spatial-temporal-frequency (STF) features, enabling the recognition and classification of radar targets. The simulation results verify the effectiveness of the proposed method in the low signal noise ratio scenarios. Specifically, it achieved 5% accuracy improvement at
$\mathbf{0}\ dB$
, compared to suboptimal performance method. Moreover, the adjacency matrix heatmap is introduced to highlight the similarity between radars.