基于多模式仿真数据协同迁移的轴承故障辨识  

Bearing Fault Identification Based on Multi-modal Simulation Data Co-migration

作　　者：刘小峰[1] 亢莹莹 柏林[1] 陈兵奎[1] LIU Xiaofeng;KANG Yingying;BO Lin;CHEN Bingkui(State Key Laboratory of Mechanical Transmission for Advanced Equipment(Chongqing University),Shapingba District,Chongqing 400044,China)

机构地区：[1]高端装备机械传动全国重点实验室(重庆大学),重庆市沙坪坝区400044

出　　处：《中国电机工程学报》2024年第16期6632-6643,I0028,共13页PROCEEDINGS OF THE CHINESE SOCIETY FOR ELECTRICAL ENGINEERING

基　　金：国家科技重大专项(J2019-IV-0001-0068);国家自然科学基金项目(52175077)。

摘　　要：针对目前迁移诊断算法对训练样本规模与质量的依赖性,以及在特殊工况环境中进行轴承故障数据采集与标注的困难性问题,提出基于多模式仿真数据协同迁移的轴承故障诊断方法。首先,采用实际工况参数嵌入的轴承故障动力学模型生成各类故障模拟数据,解决实际故障样本不足与标签缺失问题。然后,根据在仿真数据可迁移模式分析基础上建立多个子源域,引入几何统计联合对齐法对各子源域进行无监督迭代迁移,克服了单一模式迁移信息不足与跨域特征差异度过大引起的负迁移问题。最后,采用优化模糊积分决策融合方法对迁移迭代中的多模式特征伪标签进行协同分配,逐步提高目标域标签的可信度与迁移模型的域适应能力。试验结果表明,该文所提方法以故障仿真数据为驱动,无需实测标签数据的引导迁移,就可实现轴承各类故障的准确辨识。该方法对工况环境变化及目标域样本大小具有较好鲁棒性,在非完备数据支撑的高端轴承故障诊断领域具有较好应用前景。In view of the dependence of the current migration diagnosis algorithm on the size and quality of training samples and the difficulty of collecting and labeling bearing fault data in the special working condition environment,a bearing fault diagnosis method based on the co-migration of multi-mode simulation data is proposed.First,the bearing fault dynamics model embedded with actual working conditions parameters is used to generate various fault simulation signals,which solves the problem of insufficient actual fault samples and missing labels.Then,multiple sub-source domains are established based on the analysis of the migratable modes of the simulation data,and the unsupervised iterative migration of each sub-source domain is introduced by the geometric statistical joint alignment method,which overcomes the negative migration problems caused by insufficient information of single mode migration and excessive differences in cross-domain features.Finally,an optimized fuzzy integral decision fusion method is used to collaboratively assign the pseudo-labels of multi-mode features in the migration iterations to gradually improve the credibility of the target domain labels and the domain adaptation capability of the migration model.The experimental results show that the proposed method is driven by fault simulation data and can achieve the accurate identification of various bearing faults without the guided migration of the measured label data.The method is robust to changes in working conditions and target domain sample size and has good application prospects in the field of high-end bearing fault diagnosis supported by non-complete data.

关 键 词：故障动力学建模 协同迁移 几何统计联合对齐 模糊积分 迁移诊断 

分 类 号：TH17[机械工程—机械制造及自动化]