Fault Diagnosis Methodologies For Rotating Machinery With Incomplete Data In Case Of Cross-domain

Fault diagnosis technology is an important part of manufacturing equipment health management,maintenance,and life prediction.Integrating new generation artificial intelligence technology and fault diagnosis technology can deeply mine the health status information from the rotating machinery,and can help achieve fault diagnosis,and following repair management of manufacturing equipment.The research on datadriven methods based on deep learning is currently a hot topic in the field of intelligent fault diagnosis.Due to the limitations of data acquisition,the following limitations still exist in relevant research:(1)Lack of labeled data within the same fault category,which refers to incomplete intra-class data.(2)Unequal quantity of fault data and health data,which refers to incomplete inter-class data.(3)Distribution difference between target domain and source domain due to cross working conditions and devices.The above situation may lead to issues such as unclear objective functions and poor generalization ability of the fault diagnosis model based on deep learning.Therefore,this article combines the methods of deep learning and transfer learning,taking bearings,gearboxes,and motors in rotating machinery as research objects,and conducts systematic research on the challenges of incomplete data in cross domain situations mentioned above.The main content is as follows:(1)To address the inability of utilizing a large amount of unlabeled data under incomplete intra-class data and fault feature shift caused by cross working conditions,a triplet metric loss guided ladder-shaped semi-supervised convolutional neural network model guided by is proposed from the perspectives of network structure and feature space.Supervised learning and unsupervised learning are compatible by constructing a ladder-shaped semi-supervised network structure,and both fault information extraction paths are provided for labeled and unlabeled data,respectively.Local connections of convolutional neural networks are used to instead of full connections,and parameter interactions are carried out in both labeled and unlabeled data paths to reduce model parameters and prevent overfitting of the model.Triplet metric loss is introduced in the embedding feature space of labeled data paths to measure the distance.Thus,the inter-class decision boundaries with larger intervals can be obtained by narrowing the distance between samples of the same class and pushing the distance between samples of different classes,and the impact of fault feature shift under different working conditions can be reduced.Finally,two engineering cases were used to validate the effectiveness and applicability of the proposed method in the case of incomplete intra-class data under cross working conditions.(2)The incomplete inter-class data cannot afford to capture potential patterns within the fault category,to address this problem,a dual branch fault diagnosis method based on spatial variable scale pooling is proposed.A dual branch parallel model consisting of VGG16 convolution and Transformer Encoder is constructed,and a feature fusion module for semantic alignment of dual branch features is designed to simultaneously extract and fuse local and global fault features,which can enhance the fault representation learning ability.The pooling layer between sub-modules in the convolutional branch is improved by spatial variable scaling pooling,which prevents the loss of fault information.Finally,two engineering cases were used to validate the effectiveness and applicability of the proposed method in the case of incomplete interclass data under constant working conditions.(3)In the case of incomplete inter-class data under cross working conditions,additional consideration should be given to the fault features shifting affected by working conditions,and a multi-scale feature fusion intelligent diagnosis model based on dual attention mechanism is proposed.A dual branch parallel model consisting of spatial multi-scale convolution and Transformer Encoder was constructed to obtain full resolution fault features of the input data,which can not only supplement fault information at different scales,but also can capture the fault frequency bands shift across working conditions.A feature fusion module based on dual attention mechanism with randomness is designed to adaptively obtain important information related to the target task,while ignoring the impact of working condition information on diagnostic results.Finally,two engineering cases were used to validate the effectiveness and applicability of the proposed method in the case of incomplete inter-class data across working conditions.(4)In the case of incomplete inter-class data across device,the feature present more significant distribution differences comparing with the across working conditions.To address the balance between the computational resources and diagnostic accuracy of large-scale deep learning models,a lightweight transfer learning diagnostic model with distillation is proposed.To learn domain invariant features related to fault classification in the source and target domain data,a deep teacher model and a student model are constructed based on variable scale residual network,respectively.A knowledge distillation framework introducing a temperature factor is constructed to transmit the fault information learned by large teacher models in the source domain,which can reduce the computational and parameter burden.A multi-kernel domain adaptation method is utilized to obtain the feature probability distribution distance of fault features in the source and target domains in the regenerated Hilbert space,and invariant features between domains can be learned by minimizing the distribution distance.Finally,the effectiveness and applicability of the proposed method are verified through two engineering cases,which including cross device types and the case from laboratory equipment to real-world equipment.