Research On Fault Diagnosis Method Based On Domain Adaptive Network

In modern industry and intelligent manufacturing,rotating machinery gradually occupies an increasingly important position.Real-time monitoring of the working status of rotating machinery can not only avoid major disasters,but also bring obvious economic benefits.Rolling bearing is one of the most used key parts in various rotating machinery and equipment,and it is also one of the most easily damaged parts.The current state-of-the-art research on rolling bearings is based on cross-domain bearing fault diagnosis,but the differences between domains make the diagnosis model unable to perform well from one domain to another.The biggest challenge in cross-domain fault diagnosis is how to improve similar distributions between domains.Domain adaptation solves the problem of fault diagnosis based on cross-domain well.Although the current image-based domain adaptive fault diagnosis can reduce the distribution difference between the source domain and the target domain,most of these methods use stronger semantic features,high-level features,ignoring some fine-grained low-level feature information,such as some texture,edge information of the image.In addition,among existing fault diagnosis methods,most of them utilize deep networks to achieve fault classification,ignoring the effect of network depth on feature transferability.Focusing on the above problems in the fault diagnosis of rolling bearings,this thesis studies the fault diagnosis in the case of domain self-adaptation and domain self-adaptive fault diagnosis in multi-source domain.The innovations and main contents of this thesis are as follows:(1)In order to solve the problem of insufficient feature extraction in cross-domain fault diagnosis,a cross-domain fault diagnosis method based on domain-specific attention is proposed.This method captures the low-level features of the data by incorporating channel attention and spatial attention on the convolutional layers.The feature distribution difference between the source and target domains is measured by using MMD,and the feature correction module is used to perform feature correction on the captured low-level and high-level features to achieve similar distributions in the source and target domains.Experimental results show that this method has advantages in cross-domain fault diagnosis.(2)Inorder to solve the problem that the feature transferability will gradually deteriorate with the depth of the network,a cross-domain fault diagnosis method based on MLP-Mixer network is proposed.This method uses MLP-Mixer network instead of Resnet50,which simplifies the network structure.A more comprehensive feature expression is obtained by replacing the attention mechanism by channel mixing and spatial position mixing of image feature tensors.In addition,the feature distribution difference between the source domain and the target domain is further reduced by the improved feature correction module,and the transferability of the domain is enhanced.The experimental results show that the method effectively enhances the transferability of the domain and improves the diagnostic accuracy.(3)A cross-domain fault diagnosis method based on multi-source domain is proposed to solve the problem of insufficient information in single-source domain and insufficient knowledge transfer.The method firstly performs common feature extraction on the shared basic feature extraction networks of different source domains,and then designs specific feature extraction networks and classification networks for each source domain.In addition,after a specific feature extraction network,the feature correction module is used to reduce the distribution difference of each pair of source domain and target domain,further reduce the inconsistency of different classifiers on the target domain sample prediction and improve the fault diagnosis accuracy.Experimental results show that this method has advantages in multi-source fault diagnosis.