Modulation Recognition of Radar Signals Based on Multimodal Contrastive Learning

Deep learning has been extensively used in radar signal modulation recognition, leading to significant improvements in accuracy. For supervised methods, the recognition performance mainly depends on the quality of large-scale labeled data. However, data annotation is usually expensive and time-consuming. The acquisition of high-quality labeled data poses a significant challenge. To address this issue, this paper proposes a radar signal modulation recognition method based on multimodal contrastive learning (RS-MCL). First, we obtain the feature of radar signal by performing pre-training based on contrastive learning with unlabeled multimodal radar signals. Then, the pre-trained encoder is fine-tuned along with a randomly initialized classifier to finish the recognition task, where only a small number of labeled samples are fed. Given the characteristics of multimodal inputs, two distinct attention mechanisms are incorporated in the encoder to effectively extract features from both the time-domain signal and time-frequency image. Experimental results demonstrate the superiority and stability of the proposed method across most of signal-to-noise ratio (SNR) conditions, even when utilizing only 1% of the labeled samples.