基于多源域深度域自适应的跨工况滚动轴承故障诊断

Fault Diagnosis of Cross-Working Rolling Bearing Based on Multi-source Depth Domain Self-adaptation

在实际生产中,不同工况下的源域数据与目标域数据分布差异大且含标签的故障样本量少,单源域迁移故障诊断无法同时利用多个源域提供的诊断信息,会出现负迁移与模型泛化能力差的问题。针对此情况,提出一种基于多源域深度域自适应模型的滚动轴承故障诊断方法。将多个源域与目标域的原始一维时域信号输入到模型中的共享特征提取网络中,提取所有域的域不变特征;利用私有特征提取网络分别匹配每个源域与目标域的特征空间分布,结合最大均方差异(MMSD)与局部最大均值差异(LMMD)设计新型损失函数——局部最大均方差异(LMMSD),减小每对源域与目标域之间的数据特征分布差异,同时,使用领域判别器损失进一步增加域混淆;最后,根据LMMSD损失获得不同源域相对于目标域的权重,将多个源分类器与相应的权重相结合,对设备状态进行综合诊断。在2个公开变工况滚动轴承故障数据集上进行实验验证,结果表明:所提方法相比其他方法具有更高的诊断精度与泛化能力。

In actual production,the difference in the distribution of source domain data and target domain data is large under different working conditions,and the number of fault samples containing labels is small,the single source domain migration fault diagnosis cannot utilize the diagnostic information provided by multiple source domains simultaneously,there are some problems of negative migration and poor model generalization ability.In view of this situation,a rolling bearing fault diagnosis method based on multi-source domain deep domain self-adaptation model was proposed.The original 1D time domain signals of multiple source and target domains were input into the shared feature extraction network of the model,and the domain invariant features of all domains were extracted.A private feature extraction network was used to match the feature spatial distribution of each source domain and target domain respectively.A new loss function of local maximum mean square discrepancy(LMMSD)was designed by combining maximum mean square discrepancy(MMSD)and local maximum mean discrepancy(LMMD),to reduce the difference in data feature distribution between each pair of source domain and target domain.Meanwhile,the loss of domain discriminator was used to further increase domain confusion.Finally,the weights of different source domains relative to target domains were obtained according to LMMSD losses,and multiple source classifiers were combined with corresponding weights to carry out comprehensive diagnosis of device status.The experimental results on two publicly available working conditions rolling bearing fault datasets show that the proposed method has higher diagnostic accuracy and generalization ability than other methods.