基于自监督对比学习的低信噪比频谱感知算法
Self-Supervised Contrastive Learning for Low-SNR Spectrum Sensing
胡博程 1高凤娇 1张敏1
作者信息
- 1. 辽宁工程技术大学电子与信息工程学院，辽宁葫芦岛，125105
- 折叠
摘要
针对认知无线电中低信噪比下传统频谱感知算法检测概率急剧下降、深度学习方法依赖大量标注样本的问题，提出一种融合自监督对比学习与深度残差收缩网络的频谱感知算法。设计SSCL-DRSN两阶段训练框架：预训练阶段基于BYOL自监督学习，利用海量无标签频谱样本学习信号通用表征；微调阶段采用深度残差收缩网络作为感知骨干，通过逐通道自适应软阈值机制抑制噪声、保留信号特征，仅需少量标签样本完成分类器训练。基于BPSK调制信号进行蒙特卡洛仿真验证。仿真表明，信噪比为-12dB时检测概率答0.80，较能量检测提升51.6个百分点；仅用10%标签数据，检测概率即达0.965，达到全监督CNN使用全量标签时98.3%的性能水平；推理延时2.3ms，参数量320万。SSCL-DRSN通过自监督预训练与自适应软阈值机制的融合，在标注数据效率、低信噪比检测性能和推理延迟之间取得良好平衡。
Abstract
To address the sharp decline in detection probability of conventional spectrum sensing algorithms under low signal-to-noise ratio (SNR) and the heavy reliance of deep learning methods on large-scale labeled datasets, a spectrum sensing algorithm integrating self-supervised contrastive learning with a deep residual shrinkage network is proposed.A two-stage SSCL-DRSN training framework is designed. In the pretraining stage, the BYOL self-supervised learning framework is adopted to learn universal signal representations from massive unlabeled spectrum samples. In the fine-tuning stage, a deep residual shrinkage network serves as the sensing backbone, where a channel-wise adaptive soft-thresholding mechanism suppresses noise while preserving signal features, requiring only a small amount of labeled data. Monte Carlo simulations are conducted based on BPSK modulated signals.Simulation results demonstrate that at an SNR of ?12 dB, the proposed algorithm achieves a detection probability of 0.80, representing an improvement of 51.6 percentage points over energy detection. Using only 10% labeled data for fine-tuning, the detection probability reaches 0.965 at an SNR of ?10 dB, attaining 98.3% of the performance level of a fully supervised CNN. The inference latency is 2.3 ms, and the model has 3.2 million parameters.SSCL-DRSN achieves a favorable balance among labeling data efficiency, low-SNR detection performance, and inference latency through the organic integration of self-supervised pretraining and adaptive soft-thresholding, providing a viable solution for practical deployment of spectrum sensing modules in cognitive radio systems.关键词
认知无线电/频谱感知/自监督对比学习/残差收缩网络Key words
cognitive radio/spectrum sensing/self-supervised contrastive learning/deep residual shrinkage network引用本文复制引用
胡博程,高凤娇,张敏.基于自监督对比学习的低信噪比频谱感知算法[EB/OL].(2026-06-23)[2026-06-25].http://www.paper.edu.cn/releasepaper/content/202606-62.学科分类
无线电设备、电信设备/无线通信/通信
