改进抗噪1D-CNN的旋转车轮动平衡状态监测  

Dynamic Balance State Monitoring of Rotating Wheels Based on Improved Noise Resistant 1D-CNN

作　　者：周海超[1] 关浩东 王国林[1] 张宇 赵春来 ZHOU Haichao;GUAN Haodong;WANG Guolin;ZHANG Yu;ZHAO Chunlai(School of Automotive and Traffic Engineering,Jiangsu University Zhenjiang,212013,China;Forward Technical Research Institute,Dongfeng Motor Co.,Ltd.Wuhan,430056,China)

机构地区：[1]江苏大学汽车与交通工程学院,镇江212013 [2]东风汽车集团有限公司前瞻技术研究院,武汉430056

出　　处：《振动.测试与诊断》2025年第2期309-315,412,413,共9页Journal of Vibration,Measurement & Diagnosis

基　　金：国家自然科学基金面上资助项目(52072156,52272366)。

摘　　要：针对实车旋转车轮动平衡状态难以实时监测及预判的问题,提出了一种融合注意力机制的抗噪一维卷积神经网络(noise resistant 1D convolutional neural network,简称NRCNN)的旋转车轮动平衡健康状态监测方法。首先,构建NRCNN模型,以在实车车轮上添加3种不同质量平衡块的方式获得3种不同速度下对应的旋转车轮动不平衡状态下的振动信息;其次,以高斯白噪声为噪声输入,对所测旋转车轮不同动平衡状态的振动信息进行处理,获得试验样本数据,并用其进行模型训练;然后,综合运用卷积运算机制和特征变换进行t分布随机邻域嵌入(t-distributed stochastic neighbor embedding,简称t-SNE)可视化显示,实现对不同动平衡状态的分类输出。结果表明,在不同信噪比的工况下,所提出的改进NRCNN模型旋转车轮的动平衡状态监测方法相比于传统一维卷积神经网络(1D convolutional neural network,简称1D-CNN)模型,展现出更高的诊断准确性,最高可达到99.95%。Aiming to the difficulties for monitoring and determining the dynamic balance state of the rotating wheel of the real vehicle in real time,a dynamic balance monitoring method of the rotating wheel dynamic balance is proposed with the attention mechanism of the noise resistant 1D convolutional neural network(NRCNN).First,the vibration information under the dynamic imbalance state of the corresponding wheel movement at three different speeds is obtained by adding three different mass balance blocks to the real wheel.Secondly,the Gaussian white noise is used as the noise input,and the vibration information of the measured rotating wheel in different dynamic equilibrium states is processed to obtain the test sample data.Then,the t-distributed stochastic neighbor embedding(t-SNE) visualization display is carried out by comprehensively using the convolutional operation mechanism and feature transformation to realize the classified output of different dynamic equilibrium states.The results show that,compared with the traditional 1D convolutional neural network(1D-CNN) model,the NRCNN rotating wheel dynamic balance state monitoring method shows better diagnostic accuracy than the 1D-CNN model,which can reach 99.95% under different signal-to-noise ratios.

关 键 词：卷积神经网络 注意力机制 车轮动平衡 状态监测 高斯白噪声 

分 类 号：TH17[机械工程—机械制造及自动化]