基于循环生成式对抗网络的一维时变信号自适应去噪研究  

Research on Adaptive Denoising of 1D Time-varying Signal Based on CycleGAN

作　　者：董骏捷 唐建 周然之 杨超越 Dong Junjie;Tang Jian;Zhou Ranzhi;Yang Chaoyue(PLA Army Engineering Universit,Nanjing 210007,China)

机构地区：[1]陆军工程大学野战工程学院,南京210007

出　　处：《机电工程技术》2021年第5期10-12,17,共4页Mechanical & Electrical Engineering Technology

基　　金：国家自然科学基金资助项目(编号:51705531);江苏省自然科学基金资助项目(编号:BK20150724)。

摘　　要：针对传统去噪方法自适应差、对低信噪比时变信号去噪能力不足的问题,提出了一种基于循环生成式对抗网络CycleGAN的信号去噪方法。将广泛用于二维图像数据风格迁移的CycleGAN改进为适用于时序信号的一维CycleGAN,通过含噪信号和无噪信号两个数据集的循环对抗训练,得到信号从含噪空间到无噪空间的端到端最佳映射,从而获得具备自适应降噪功能的去噪模型。经过6组添加了不同信噪比的高斯白噪声的含噪信号集的测试实验,结果表明,该方法对于低信噪比的含噪时变信号具有优越的去噪能力,在信噪比和均方误差这两项指标的评价上都显著优于传统方法。Aiming at the problems of poor adaptive denoising of traditional denoising methods and insufficient denoising ability for low signal-tonoise ratio time-varying signals,a signal denoising method based on CycleGAN was proposed.CycleGAN,which had been successfully applied to the style transfer of two-dimensional image data,was improved to a one-dimensional CycleGAN suitable for time series signals.Through the cyclical confrontation training by two data sets of noisy signal and no-noise signal,the best end-to-end mapping of the signal from the noisy space to the no-noise space was obtained,thereby obtaining a denoising model with adaptive noise reduction function.After six groups of test experiments on noisy signal sets with Gaussian white noise with different signal-to-noise ratios added,the results show that this method has superior denoising ability for high-noise time-varying signals.This method has superior denoising ability for Noisy time-varying signal with low SNR,and is significantly better than traditional methods in the evaluation of the two indicators of signal-to-noise ratio and mean square error.

关 键 词：去噪 时变信号 循环生成式对抗网络 映射 自适应 

分 类 号：TN911.6[电子电信—通信与信息系统]