摘要
针对转子裂纹故障智能诊断模型研究中存在的不足之处,如需要大量数据样本作为支撑、数据复用性差、无法识别裂纹的扩展情况等,提出了一种基于数据驱动的域适应迁移学习模型。该模型基于三元组损失和降噪自编码网络,基于对抗训练策略提取域无关裂纹故障特征,实现不同领域特征的全局对齐。使用三元组损失约束故障特征,实现不同领域故障特征的类级特征对齐。模型以裂纹转子运行数据为输入,预测裂纹的扩展阶段。对模型的跨工况特征迁移效果测试结果显示,10个不同跨工况特征迁移任务的平均预测准确率为91.3%。相较于其他经典的迁移学习模型,该模型能够提取更有效的域无关裂纹故障特征,具有更强的特征迁移泛化效果。
Aiming at the deficiencies in the research of the intelligent diagnosis model for rotor crack faults,such as the requirement for a large number of data samples as the support,poor data re-usability,and the impossibility to identify the propagation situation of crack,a data-driven domain adaptation transfer learning model was proposed.In this model,based on triplet loss and denoising autoencoder network,domain-invariant crack fault features were extracted using adversarial training strategy to achieve global alignment of features from different domains.The triplet loss was employed to constrain the extracted fault features and realize class-level feature alignment across different domains.With the vibration signal of cracked rotor as input,the propagation stage of crack was predicted.Finally,the cross-operating condition feature transfer performance of the model was tested.The results show that the average prediction accuracy of 10 different cross-operating condition feature transfer tasks reaches 91.3%.Comparing with other classical transfer learning models,the proposed model can extract more effective domain-invariant crack fault features and exhibit stronger feature transfer generalization capabilities.
作者
陈志昊
赵文强
王正伟
周军
石生超
李富才
CHEN Zhihao;ZHAO Wenqiang;WANG Zhengwei;ZHOU Jun;SHI Shengchao;LI Fucai(The State Key Laboratory of Mechanical System and Vibration,Shanghai Jiaotong University,Shanghai 200240,China;State Grid Qinghai Electric Power Company,Qinghai 810000,China)
出处
《噪声与振动控制》
北大核心
2025年第3期105-112,共8页
Noise and Vibration Control
基金
国网青海省电力公司科技资助项目(522807230005)。
关键词
故障诊断
裂纹转子
迁移学习
域适应理论
三元组损失
fault diagnosis
cracked rotor
transfer learning
domain adaptation theory
triplet loss
作者简介
陈志昊(1997-),男,河南省洛阳市人,博士研究生,专业方向为旋转机械智能故障诊断模型研究。E-mail:chen_zhihao@sjtu.edu.cn;通信作者:李富才,男,博士生导师,专业方向为结构健康监测、机械故障诊断。E-mail:fcli@sjtu.edu.cn。
