一种自适应残差卷积自编码网络及其故障诊断应用  

Adaptive residual convolutional auto-encoder network and its fault diagnosis application

作　　者：潘天成[1] 陈龙[1] 蒲春雷[1] 陈志强[1] PAN Tiancheng;CHEN Long;PU Chunlei;CHEN Zhiqiang(MCC Huatian Engineering&Technology Co.,Ltd.,Nanjing 210019,China)

机构地区：[1]中冶华天工程技术有限公司,江苏南京210019

出　　处：《机电工程》2025年第3期529-538,共10页Journal of Mechanical & Electrical Engineering

基　　金：安徽省高校自然科学研究重点项目(2022AH050315)。

摘　　要：针对传统卷积自编码器(CAE)会将不同故障产生的相似信号进行相同的非线性变换,导致故障诊断准确率下降的问题,提出了一种自适应残差卷积自编码网络(ARCAE),并将其应用于滚动轴承故障诊断中。首先,在残差模块的基础上,引入了自适应参数化修正线性单元(APReLU),建立了自适应残差模块(ARM),ARM可以对相似的输入特征进行自适应非线性变换,避免了特征的错误识别;其次,在CAE中嵌入多级ARM,构建了ARCAE,增加了CAE的深度,提取了更具鉴别性的深层次特征,同时有效防止了网络加深而造成的性能退化;最后,基于ARCAE建立了针对一维信号的故障诊断新方法,将其应用于无监督滚动轴承故障诊断中,并通过两个不同类型的实验,对上述方法的有效性进行了验证。研究结果表明:在恒定转速工况下,ARCAE的诊断准确率最高,平均准确率达到了97.05%,且标准差仅为0.007,远低于其他几种传统CAE网络;在变转速工况下,ARCAE模型诊断准确率仍然是最高的,平均准确率达到了93.25%,由此说明ARCAE具有较高的特征提取能力和分类准确率;此外,变转速工况下,由于转速变化导致不同状态的振动信号特征差异变大,诊断难度加大,但与其他几种传统CAE网络相比,ARCAE诊断准确率下降最少,仅为5.37%,说明ARCAE具有更强的鲁棒性和稳定性。In response to the problem of traditional convolutional auto-encoder(CAE)performing the same nonlinear transformation on similar signals generated by different faults,resulting in a decrease in fault diagnosis accuracy,an adaptive residual convolutional auto-encoder network(ARCAE)was proposed and applied to the fault diagnosis of rolling bearings.Firstly,based on the residual module,an adaptive parametric rectifier linear unit(APReLU)was introduced to establish an adaptive residual module(ARM).Adaptive nonlinear transformation was carried out by ARM on similar input features to avoid wrong recognition of features.Secondly,multi-level ARM was embedded in CAE to build ARCAE,the depth of CAE was increased,more discriminative deep-seated features were extracted,and performance degradation was effectively prevented caused by network deepening.Finally,a new fault diagnosis method for one-dimensional signals based on ARCAE was established and applied to unsupervised rolling bearing fault diagnosis,validating through two different types of experiments.The research results show that under constant speed conditions,ARCAE has the highest diagnostic accuracy,with an average accuracy of 97.05%and a standard deviation of only 0.007,far lower than other traditional CAE networks.Under variable speed conditions,the ARCAE model still has the highest diagnostic accuracy,with an average accuracy of 93.25%.This indicates that ARCAE has high feature extraction capability and classification accuracy.In addition,under variable speed conditions,the differences in vibration signal characteristics between different states due to changes in speed increase the difficulty of diagnosis.However,comparing with other traditional CAE networks,ARCAE has the least decrease in diagnostic accuracy,only 5.37%,indicating that ARCAE has stronger robustness and stability.

关 键 词：滚动轴承 自适应残差卷积自编码网络 自适应参数化修正线性单元 自适应残差模块 无监督故障诊断 特征提取 

分 类 号：TH133.3[机械工程—机械制造及自动化]