基于深度自编码器的振动信号盲去噪方法  

A blind denoising method of vibration signals based on a deep autoencoder

作　　者：万若青 张纯[1] 江汇强 黎寅斌 WAN Ruoqing;ZHANG Chun;JIANG Huiqiang;LI Yinbin(School of Infrastructure Engineering,Nanchang University,Nanchang 330031,China)

机构地区：[1]南昌大学工程建设学院,南昌330031

出　　处：《振动与冲击》2023年第12期118-125,共8页Journal of Vibration and Shock

基　　金：国家自然科学基金(51469016,51968047,52268050);江西省自然科学基金(20202BAB204029);江西省学位与研究生教育教学改革(JXYJG-2019-018)。

摘　　要：为减少传统振动信号去噪方法对信号时、频域先验信息的依赖性,提出了一种基于深度自编码器的振动信号盲去噪方法。在缺少干净信号作为神经网络训练目标的情况下,使用邻近采样及扩展的策略,从原始信号中构造去噪深度神经网络的训练样本对,通过自监督学习得到能对原始信号有效降噪的深度神经网络;并提出适用性评价指标来指导在实际工程应用时信号采样频率的设置。对仿真信号和实测信号的去噪分析表明该方法不依赖于真实信号的先验信息,且对于稳态信号和非稳态信号都有良好的自适应去噪效果。In order to reduce the dependence of traditional vibration signal denoising methods on the prior information in time or frequency domain,a blind vibration signal denoising method based on a deep autoencoder was proposed.In the absence of clean signals as the training target of the neural network,the adjacent sampling and expansion strategy was used to construct the training sample pair of the denoising deep neural network from the original signal.The deep neural network that can effectively denoise the original signals was obtained through self-supervised learning.And the applicability evaluation index was proposed to guide the setting of signal sampling frequency in practical engineering application.The denoising analysis of simulation signals and measured signals show that the proposed method does not depend on the prior information of real signals and has great adaptive denoising effect for both steady and unsteady signals.

关 键 词：振动信号 深度学习 去噪自编码器 盲去噪方法 

分 类 号：TB123[理学—工程力学]