基于相似性度量迁移学习的轴承故障诊断

A bearing fault diagnosis based on similarity measurement for transfer learning

滚动轴承工况多变,受负荷、转速等因素影响,故障信号的特征分布偏移会显著降低故障诊断模型的泛化能力。针对此问题,提出一种基于相似性度量迁移学习的轴承故障诊断方法。将迁移学习和相似性度量的思想结合,通过相关对齐损失计算变工况故障特征之间的相关性,最小化源域和目标域特征之间的分布差异。同时最大化输入特征与中心特征的相似性,利用目标域预测标签中包含的故障分类信息,提高故障特征聚类的准确性,来增加高相关性特征对模型的贡献度,减小非相关特征的影响。最后利用学习到的特征实现故障分类。在CWRU和自搭建试验平台上进行了对比试验,证明了所述方法能够更加准确地分类故障信号,更好解决不同工况下由特征分布偏移带来的故障诊断难点问题。

According to the variable working conditions of rolling bearings due to the influence of load,speed,and other factors,the feature distribution deviation of fault signals significantly reduce the generalization ability of fault diagnosis model.To solve this problem,a bearing fault diagnosis method based on similarity measurement for transfer learning was proposed.The idea of transfer learning and similarity measurement were combined.The correlation between fault features under variable conditions was calculated through correlation alignment loss,and the distribution difference between source domain and target domain features was minimized.At the same time,the similarity between the input feature and the central feature was maximized,and the fault classification information contained in the label predicted by the target domain was used to improve the accuracy of fault feature clustering,so as to increase the contribution of highly correlated features to the model and reduce the influence of non-correlated features.Finally,the features learned were used to implement fault classification.Comparison experiments on CWRU and a self-built experimental platform prove that the proposed method can classify fault signals more accurately and solve the difficult problems of fault diagnosis caused by feature distribution deviation under different working conditions.