Contrastive Language-Signal Prediction for Automatic Modulation Recognition

Abstract

Automatic Modulation Recognition (AMR) enables intelligent communication and is a critical component of wireless communication systems. Deep learning-based AMR approaches have made significant strides in recent years. These approaches involve inputting signals in the form of images or embeddings into a network, which maps them into high-dimensional feature vectors for subsequent classification. However, radio frequency (RF) signals exhibit significant differences within the same class due to noise or wireless channels. Performing classification based on high-dimensional features may be challenging in capturing robust discriminative features, thereby compromising the model’s generalization ability. To address this limitation, we introduce a novel framework named CLASP, which incorporates language models through contrastive learning, coupling AMR with human language priors to extract robust discriminative features between different categories. Additionally, we treat the prediction of SNR levels as a subtask to acquire auxiliary priors that represent the impact of noise. Extensive results on widely-used datasets demonstrate that CLASP achieves state-of-the-art (SOTA) performance compared to other baselines. As a framework, CLASP exhibits universality and demonstrates superior performance compared to the linear-probe approach across different backbones.