EET-MoCo: An Efficient Embedding Transformer With Momentum Contrast Learning for Automatic Modulation Recognition

Automatic modulation recognition (AMR) using deep learning (DL) techniques has attracted great interest in radio and spectrum management. However, as most signals in practical noncooperative scenarios are unlabeled, DL methods may struggle to achieve satisfactory performance. In this paper, we propose the efficient embedding Transformer with Momentum Contrast (EET-MoCo), which is a self-supervised learning framework designed for AMR. Specifically, we develop a Transformer-based lightweight backbone integrated with the efficient embedding module (EEM). The EEM captures long-term correlations in modulated signals through an expanded receptive field, allowing robust local feature extraction and enhanced noise resistance. During the momentum contrastive pre-training stage, we formulate a data augmentation strategy grounded in the propagation characteristics of wireless signals. This tailored augmentation strategy generates more realistic sample pairs, enabling the proposed EET-MoCo to consistently identify dynamic features in various modulated signals and reduce the reliance on labeled samples. Comprehensive simulation experiments using three public datasets illustrate the beneficial effects of EET-MoCo. Compared with other popular methods, the proposed framework exhibits superior performance with fewer labeled samples and excellent computational efficiency. These results highlight its potential in real-world wireless communications.