强噪声下基于ACYCBD-MTF-MobileViT的轴承故障诊断研究  

A study on bearing fault diagnosis based on ACYCBD-MTF-MobileViT under strong noise

作　　者：刘杰[1] 谭玉涛 杨娜 LIU Jie;TAN Yutao;YANG Na(School of Mechanical Engineering,Shenyang University of Technology,Shenyang 110870,China)

机构地区：[1]沈阳工业大学机械工程学院,沈阳110870

出　　处：《振动与冲击》2024年第24期34-47,共14页Journal of Vibration and Shock

基　　金：辽宁省教育厅(LQGD2020016)。

摘　　要：针对小样本强噪声环境下,传`统深度学习模型抗噪性差,模型训练不充分等问题,提出一种基于自适应最大二阶循环平稳盲解卷积(adaptive maximum second-order cyclostationarity blind deconvolution, ACYCBD)结合马尔可夫变迁场(Markov transition field, MTF)与MobileViT的滚动轴承故障诊断方法。首先,通过参数自适应的CYCBD算法增强强噪声背景下轴承故障的冲击信号,降低强背景噪声的影响,然后,采用MTF将预处理后的一维轴承振动信号转变为具有时间关联性的二维特征图像;最后,将MTF图像输入MobileViT网络中进行训练,得到故障诊断结果,运用东南大学齿轮箱数据集和沈阳工业大学实验室滚动轴承数据集验证所提方法在小样本强噪声条件下的故障识别准确率。结果表明:在小样本强噪声条件下,ACYCBD处理后的数据,训练的模型具有更高的准确率,相较于其他数据预处理方法最大相关峭度解卷积、变分模态分解、集合经验模态分解准确率分别提高了1.73、1.99、2.20个百分点,利用MTF进行模态转换后相较于格拉姆角场、连续小波变换、RP准确率分别高出了2.59、3.12、2.72个百分点;与其他深度学习模型进行对比,所提方法在上述条件下有着更高的抗干扰能力和泛化性能。Aiming at the problems of poor noise robustness and insufficient model training of traditional deep learning models in small-sample and strong noise environment,a method based on adaptive maximum second-order cyclostationarity blind deconvolution(ACYCBD)combined with Markov transition field(MTF)and MobileViT was proposed for rolling bearing fault diagnosis.Firstly,the impact signal of bearing faults under strong noise background was enhanced by a parameter-adaptive CYCBD algorithm to reduce the influence of strong background noise,then,MTF was used to transform the preprocessed one-dimensional bearing vibration signal into a two-dimensional feature image with temporal correlation,and finally,the MTF image was input into the MobileViT network for training to get fault diagnosis results,which is applied to the Southeast University Gearbox Dataset and Shenyang University of Technology Laboratory rolling bearing dataset to verify the fault identification accuracy of the proposed method in small sample strong noise conditions.Results show that,in the small sample strong noise conditions,for ACYCBD preprocessed data,the trained model has a higher accuracy.Compared to the data preprocessed by maximum correlated kurtosis deconvolution,variational mode decomposition,ensemble empirical mode decomposition,the accuracy increased by 1.73%,1.99%,and 2.2%,respectively.After using MTF for modal transformation,the accuracy is 2.59%,3.12%,and 2.72%higher than that of Gramian angular field,continuous wavelet transform and RP,respectively.Compared with other deep learning models,the method proposed in this paper has higher anti-interference ability and generalization performance under the above conditions.

关 键 词：滚动轴承 最大二阶循环平稳盲解卷积(ACYCBD) 马尔可夫变迁场(MTF) 多头自注意力机制 

分 类 号：TH133.33[机械工程—机械制造及自动化]