Page 99 - 《振动工程学报》2025年第11期
P. 99

第 38 卷第 11 期                      振 动 工 程 学 报                                      Vol. 38 No. 11
               2025 年  11 月                    Journal of Vibration Engineering                       Nov. 2025



                  面   向   不    同   标   签   与    域   配   置   的   统    一   跨   域   故   障    诊   断   方   法



                       张宇腾 , 吕宇璠 , 孔 运                1,2,3 , 陈 科 , 闫志武 , 董明明 , 褚福磊                  5
                                           1
                               1
                                                                               4
                                                                   1,4
                                                                                          1
                    (1. 北京理工大学机械与车辆学院,北京 100081; 2. 重庆大学高端装备机械传动全国重点实验室,重庆 400044;
                     3. 北京理工大学唐山研究院,河北 唐山 063015; 4. 内蒙古第一机械集团股份有限公司,内蒙古 包头 014032;
                                                5. 清华大学机械工程系,北京 100084)

              摘要:可靠的设备健康监测与故障诊断技术是保证高端装备安全高效运行的关键。基于无监督域自适应的跨域智能诊断技
              术已在跨设备、变工况等迁移诊断场景中展现出广阔的应用前景。然而,此类方法依赖域间标签关系和域配置的特定事前假
              设,致使无监督域自适应技术在实际工业故障诊断场景中的泛化性与实用性受限。针对上述问题,本文提出一种面向不同标
              签与域配置的统一跨域故障诊断方法。该方法构建一种多场景共享的预测类别混淆偏差用于指导跨域知识迁移,从而适应
              各种跨域故障诊断场景。为更准确地度量预测类别混淆偏差,提出一种基于原型相似度的故障判别方法以增强分类鲁棒性,
              从而为估计预测类别混淆偏差提供可靠的预测分布。此外,设计了一种基于标签平滑的概率校准方法进行概率正则化,以缓
              解过度自信预测导致的预测类别混淆偏差低估。行星齿轮箱传动系统数据集试验验证结果显示,所提方法在                                     4 种不同标签和
              域配置的跨域诊断场景中,平均诊断准确率达到                 98.37%,相较于前沿对比方法具有优势,充分验证了所提方法的通用性和优
              越性。

              关键词: 智能故障诊断;多场景跨域诊断;统一方法;迁移学习
                             +
              中图分类号:TH165 .3        文献标志码:A        DOI:10.16385/j.cnki.issn.1004-4523.202508021

                         Unified cross-domain fault diagnosis method towards different label

                                               and domain configurations

                                            1
                                                                              4
                                 1
                                                                  1,4
                                                                                              1
                    ZHANG Yuteng ,LYU Yufan ,KONG Yun  1,2,3 ,CHEN Ke ,YAN Zhiwu ,DONG Mingming ,CHU Fulei 5
                             (1.School of Mechanical Engineering,Beijing Institute of Technology,Beijing 100081,China;
                 2.State Key Laboratory of Mechanical Transmission for Advanced Equipment,Chongqing University,Chongqing 400044,China;
               3.Tangshan Research Institute,Beijing Institute of Technology,Tangshan 063015,China;4.Inner Mongolia First Machinery Group Co.,
                     Ltd.,Baotou 014032,China;5.Department of Mechanical Engineering,Tsinghua University,Beijing 100084,China)
              Abstract: Reliable  fault  diagnosis  is  crucial  to  ensuring  the  safe  and  efficient  operation  of  high-end  industrial  equipment.  Cross-domain
              intelligent  diagnosis  technologies  based  on  unsupervised  domain  adaptation  (UDA)  have  demonstrated  promising  application  prospects  in
              scenarios such as cross-equipment and variable working transfer diagnosis conditions. However,their effectiveness highly relies on specific
              prior assumptions regarding the inter-domain label relationships and domain configurations,which largely restricts the generalizability and
              practicality of UDA techniques in actual industrial fault diagnosis scenarios. To address the above issues,this paper proposes a unified cross-
              domain fault diagnosis framework applicable to different label and domain configurations. The proposed framework constructs a predictive class
              confusion  (PCC)  bias  shared  across  multiple  scenarios  to  guide  cross-domain  knowledge  transfer, enabling  adaptation  to  various  transfer
              diagnostic  scenarios.  To  accurately  measure  the  tendency  of  the  PCC  bias, a  prototype  similarity-based  fault  discrimination  method  is
              developed,which enhances classification robustness and provides reliable prediction distributions to estimate the PCC bias. Then,a label
              smoothing-based  probability  calibration  method  is  designed  for  probability  regularization, alleviating  the  underestimation  of  the  PCC  bias
              caused by overconfident prediction. Experimental validation results on a planetary gearbox transmission system dataset demonstrate that the
              proposed method achieves an average diagnostic accuracy of 98.37% across four cross-domain diagnostic scenarios with different label and


                  收稿日期:2025-08-11;修订日期:2025-09-27
                  基金项目:国家自然科学基金资助项目(52575094,52105108);北京市自然科学基金资助项目(3252008);中国科协青年人才
                          托举工程(2023QNRC001);北京市科协青年人才托举工程(BYESS2024294);高端装备机械传动全国重点实验
                          室开放基金资助项目(SKLMT-MSKFKT-202304);河北省自然科学基金面上项目(E2023105039);安徽省研究
                          生学术创新项目(RC2500000632)
   94   95   96   97   98   99   100   101   102   103   104