基于轻量化注意力机制的信号识别研究

打开文本图片集

中图分类号：TN98 文献标志码：A 文章编码：1672-7274（2026）01-0011-04

Lightweight Attention Mechanism-Based Signal Recognition

ZHANG Jun， WANG Baosong， SUN Zhigang， SUN Hongzhi， MIAO Lisong （Jilin Province Informationization Construction Promotion Center， Changchun 13oo33， China）

Abstract： This paper systematicallycompares two lightweight attention-based models，EdgeViTandEfficientViT， with twoclassicalconvolutional architectures，CNNand ResNet，ontheRadioML2018.0lAdataset，evaluating their performance intermsofaccuracy，convergencebehavior，modelcomplexityand inferenceeficiency.Experimental results show that EficientViT achieves the most balanced performanceamong lightweight models，reaching 76% accuracy withonly 4.80k parameters and 4.82 MFLOPs，while maintaining stable convergence and moderate inference latency.These results demonstrate that EfficientViT achieves an optimal trade-off between accuracy and efficiency， validating its potential for deployment in resource-constrained wireless systems.

Keywords：automatic modulation classification; lightweightatention; edgecomputing; wireless signal clasification

研究背景

近年来，自动调制识别（AutomaticModulationClassification，AMC）已成为频谱监测、认知无线电等无线通信领域的关键技术，广泛应用于边境车载监测、非法广播监测及“黑飞”无人机检测等任务。（剩余6307字）