基于中间桥层和相似矩阵的深度对抗迁移故障诊断方法

Fault diagnosis method of deep adversarial transfer based on middle bridge layer and similarity matrix

采用深度对抗迁移学习算法进行故障诊断时,受到领域中丰富的特征属性的影响,在领域自适应中无法充分学习可用于迁移的共有知识特征,且其在类别水平上忽略了不同类别的对齐程度的差异。针对这一问题,提出了一种基于中间桥层和相似矩阵(MB-SM)的对抗故障诊断模型,以实现对滚动轴承故障进行跨域诊断识别的目的。首先,利用改进的一维多尺度残差网络对数据的特征进行了提取;然后,引入了中间桥层和相似矩阵,完成了对共有知识特征的充分学习,降低了整体网络的数据传输难度,进一步加强了源域和目标域中同一类别内的聚类和类别之间的分离,提高了故障数据的领域适配能力;最后,采用实验室轴承数据集和美国凯斯西储大学(CWRU)数据集,对基于中间桥层和相似矩阵的模型方法进行了验证。研究结果表明:在自建实验室数据集中,采用基于中间桥层和相似矩阵的方法可以达到90.37%的平均准确率;在美国凯斯西储大学(CWRU)数据集中,也可以达到99.34%的平均准确率。相较于其他迁移学习对比模型,采用该模型方法可以获得更好的诊断性能。

In view of the existing fault diagnosis algorithms based on deep adversarial transfer learning could not fully learn the common knowledge features which could be used for transfer due to the influence of the rich feature attributes in the domain adaptation,and also the problem that the alignment degree of different categories was different at the category level in the global distribution alignment was ignored,an adversarial fault diagnosis model based on middle bridge layer and similarity matrix(MB-SM)was proposed to fully learn the common knowledge features,and realize the purpose of cross-domain diagnosis and identification of rolling bearing fault.Firstly,the improved one-dimensional multi-scale residual network was used to extract the features of the data.Then the middle bridge layer and similarity matrix were introduced to realize the full learning of the common knowledge features and reduce the difficulty of data transmission in the whole network.It further strengthened the clustering and the separation of categories in the same category in the source domain and target domain,and improved the domain adaptation ability of fault data.Finally,the proposed model method was verified by using laboratory bearing data set and Case Western Reserve University(CWRU)data set.The research results show that the average accuracy of the proposed method is 90.37%on the self-built laboratory data set and is 99.34%on the Case Western Reserve University(CWRU)data set,which can achieve better diagnostic performance than other transfer learning comparison models.