MCLRL: A Multi-Domain Contrastive Learning with Reinforcement Learning Framework for Few-Shot Modulation Recognition

TL;DR

The paper tackles AMR under data scarcity by introducing MCLRL, a framework that fuses multi-domain contrastive learning with reinforcement learning to exploit unlabeled signals for few-shot classification. It leverages three signal representations—IQ time domain, frequency domain, and constellation diagrams—and uses SAC to autonomously select augmentation strategies, guided by a multi-domain contrastive loss that combines intra- and inter-domain terms. Empirical results on RML2016.10a and Sig2019-12 show that MCLRL outperforms image-domain CL, time-series CL, and existing few-shot AMR methods, particularly in 1-shot and 5-shot scenarios, with gains of about 3–15% in high-SNR settings. The approach provides a scalable solution for robust AMR with limited labeled data, reducing labeling costs while preserving accuracy across diverse modulation schemes and channel conditions.

Abstract

With the rapid advancements in wireless communication technology, automatic modulation recognition (AMR) plays a critical role in ensuring communication security and reliability. However, numerous challenges, including higher performance demands, difficulty in data acquisition under specific scenarios, limited sample size, and low-quality labeled data, hinder its development. Few-shot learning (FSL) offers an effective solution by enabling models to achieve satisfactory performance with only a limited number of labeled samples. While most FSL techniques are applied in the field of computer vision, they are not directly applicable to wireless signal processing. This study does not propose a new FSL-specific signal model but introduces a framework called MCLRL. This framework combines multi-domain contrastive learning with reinforcement learning. Multi-domain representations of signals enhance feature richness, while integrating contrastive learning and reinforcement learning architectures enables the extraction of deep features for classification. In downstream tasks, the model achieves excellent performance using only a few samples and minimal training cycles. Experimental results show that the MCLRL framework effectively extracts key features from signals, performs well in FSL tasks, and maintains flexibility in signal model selection.

Figures (7)

• Figure 1: The heatmap-based constellation diagrams for four different categories in the RML2016.10a dataset are shown below. In these diagrams, the color darkens as the points become more concentrated.
• Figure 2: Multi-domain contrast learning MCLRL framework. The MCLRL framework can be divided into four parts: data processing, unsupervised pretraining, supervised fine-tuning, and downstream task testing.
• Figure 3: The left side of the image depicts the attention module. On the right side is the linear classifier.
• Figure 4: (a) and (b) represent the RML2016.10a dataset, indicated by the '-o-' symbol, while (c) and (d) represent the Sig2019-12 dataset, denoted by the '-*-' symbol. Taking RML2016.10a as an example, (a) shows the encoder trained in a supervised manner, where 'T_90%' indicates that the time domain encoder uses 90% of the training set samples. 'F' denotes the frequency domain encoder, while 'C' denotes the constellation diagram encoder. (b) shows the MCLRL framework in an unsupervised FSL setting, where '0.125%' represents the 1-shot sample size used for fine-tuning.
• Figure 5: The experimental results on the effectiveness of multi-representation domains and reinforcement learning in the MCLRL framework are shown. 'T' represents the time domain, 'F' represents the frequency domain, and 'C' represents the constellation diagram domain. Whether reinforcement learning is used is indicated by 'R'.