Feature-Sequence Contrastive Learning for Few-Shot Modulation Recognition

Tao Chen

doi:10.1109/tccn.2026.3705814

Feature-Sequence Contrastive Learning for Few-Shot Modulation Recognition

2026-06-22

0

PRE

AI

DOI：10.1109/tccn.2026.3705814

Original

Original request for help

Save

Abstract

En 中文

Deep learning-based modulation recognition has greatly improved performance but typically relies on large labeled datasets, making few-shot scenarios highly challenging. To address this limitation, this paper proposes a feature sequence contrastive learning (FSCL) method designed to address the performance limitations of deep learning-based modulation recognition in few-shot scenarios. The framework consists of two main stages: feature-sequence contrastive pretraining (FSCP) and sequence branch fine-tuning (SBFT). By constructing paired sequence and handcrafted feature inputs and introducing contrastive learning, the model effectively captures the complex relationships between features and sequences under unsupervised learning. Subsequently, only a small number of labeled samples are required to fine-tune the sequence branch for efficient prediction. The method not only leverages handcrafted feature information to significantly enhance recognition accuracy and robustness but also reduces the computational complexity during the inference phase. Extensive experiments across multiple datasets demonstrate that the FSCL framework achieves strong performance under extremely limited data conditions. The method shows excellent adaptability to unseen modulation types and maintains high robustness across a wide range of SNR levels, highlighting its effectiveness as a reliable solution for few-shot wireless signal recognition.

Keywords:

Deep learning

automatic modulation recognition

contrastive learning

feature-sequence pairs

Feature-Sequence Contrastive Learning for Few-Shot Modulation Recognition

Abstract

Journal

Organization