因果启发的深度域泛化旋转机械故障诊断  

Causality Inspired Deep Domain Generalization for Rotating Machinery Fault Diagnosis

作　　者：郭畅 赵志斌 张兴武[2] 刘一龙 陈雪峰[2] GUO Chang;ZHAO Zhibin;ZHANG Xingwu;LIU Yilong;CHEN Xuefeng(Shaanxi Key Laboratory of Expressway Construction Machinery,Chang'an University Xi'an,710064,China;School of Mechanical Engineering,Xi'an Jiaotong University Xi'an,710049,China)

机构地区：[1]长安大学陕西省高速公路施工机械重点实验室,西安710064 [2]西安交通大学机械工程学院,西安710049

出　　处：《振动.测试与诊断》2025年第1期154-160,206,共8页Journal of Vibration,Measurement & Diagnosis

基　　金：陕西省高速公路施工机械重点实验室(长安大学)开放基金资助项目(300102253512);国防技术基础科研资助项目(JSZL2022110A074)。

摘　　要：针对设备工况变化导致基于深度学习(deep learning,简称DL)的故障诊断性能退化的问题,提出采用因果表示网络(causal representation net,简称CRNet)用于在变工况下实现高性能故障诊断,即域泛化(domain generalization,简称DG)故障诊断。首先,假设DG的结构因果模型,并基于此模型和独立因果机制(independent causal model,简称ICM)原理,得到因果驱动的诊断需求来消除特征间的关联;其次,利用随机傅里叶特征(random Fourier features,简称RFF)将模型提取的特征映射到高维空间,再利用高维空间中的特征构造衡量特征间关联的协方差矩阵,以矩阵非对角值为目标,学习一组权重对样本加权,消除特征间的广义关联;最后,以梯度为引导,屏蔽部分高梯度特征,增强特征包含的诊断信息。锥齿轮传动实验台的实验结果表明,CRNet具备最优的DG性能。Varying working conditions of industrial equipment lead to significant degradation in the performance of deep learning(DL)-based intelligent fault diagnosis(IFD).Causal representation net(CRNet)is proposed to achieve diagnosis under varying working conditions with higher performance,i.e.,domain generalization(DG).Firstly,a structural causal model(SCM)representing DG is pre-assumed.It is then concluded,based on the SCM and the independent causal mechanism(ICM)principle,that IFD driven by causality eliminates the correlation between features.To achieve this,features are mapped into a high-dimensional space using ran⁃dom Fourier features(RFF),and covariance matrices are constructed from these high-dimensional features.Subsequently,a set of weights is obtained by optimizing the non-diagonal values of the covariance matrices.Af⁃ter weighting the samples,the general correlation between features is eliminated.Concurrently,a gradientbased feature muting training method is proposed to endow features with richer diagnosis information by muting parts of high-gradient features.Ultimately,the effective DG performance of CRNet is validated through experi⁃ments on a dataset collected from a bevel gear transmission.

关 键 词：智能诊断 深度学习 因果表示 域泛化 

分 类 号：TH165.3[机械工程—机械制造及自动化] TP183[自动化与计算机技术—控制理论与控制工程]