A Confidence-Aware Semi-Supervised Learning Framework for Automatic Modulation Recognition

Abstract

Automatic modulation recognition (AMR) is a cornerstone of dynamic spectrum allocation and cognitive radio. Despite the promise of deep learning-based methods, their reliance on large labeled datasets limits their practicality in scenarios with scarce labeled samples. To address this, we propose a confidence-aware semi-supervised learning (CASL) framework tailored for AMR. CASL adopts a two-stage approach: first, a time-warping-based data augmentation strategy is used to extract general features through contrastive learning. In the second stage, a self-adaptive threshold mechanism generates high-quality pseudo-labels while mitigating confirmation bias using class-level information. To enhance computational efficiency and performance, we design a pyramid learnable filter-based encoder that captures multi-scale features from large-scale, long-sequence signals, reducing inference time by about half compared with the self-attention-based encoder variant. Simulation results demonstrate substantial performance gains on four public datasets, achieving a 20.43% average accuracy improvement over state-of-the-art methods in label-scarce scenarios and up to 38.64% in low signal-to-noise ratio (e.g., 0 dB) conditions. These results highlight CASL’s potential for practical AMR applications in challenging environments.