Research On Bearing Fault Diagnosis Technologies Based On Domain Adaptation

Failure of rotating machinery will not only affect the normal operation of equipment,but also may bring great economic losses,and even endanger the life safety of operators,resulting in serious accidents.Bearing is a kind of widely used and extremely important rotating machinery component.The enhancement of reliability has received considerable attention due to currently precision and complexity of current mechanical equipment,which is of great significance to the fault diagnosis of bearings.For practical engineering applications,bearing working conditions are-variable due to load,speed and other factors,and the feature distribution deviation of fault signals will reduce the generalization ability of fault diagnosis model.Based on the domain adaptation theory,this paper studies the bearing fault diagnosis technologies under variable working conditions.The existing bearing fault diagnosis methods and the theoretical basis involved are introduced,and the difficulties of current intelligent fault diagnosis are summarized and analyzed.The bearing data sets used in the experiment in this paper are grouped and selected to form data sets suitable for different experimental scenarios.In order to adaptive extraction of fault features,the 2D convolution neural network model was constructed.The bearing vibration signal is taken as the input,the classification task is accomplished by the fusion of abstract features at all layers of the network.And the structure of feature extractor and fault classifier is determined by CNN structure selection experiment.Aiming at the problem of discrepancy in fault feature distribution,a bearing fault diagnosis method based on similarity measurement for domain adaptation was proposed.The ideas of domain adaptation and similarity measurement are combined,and the intraclass differences were reduced while inter-domain correlation was aligned.A domain adaptation module,CORAL,is used to narrow the distribution discrepancy between the source domain and the target domain,so that the model can effectively learn the domaininvariant features.The similarity measurement is introduced to maximize the similarity between the input feature and the central feature,and the fault classification information contained in the predict labels of the target domain is used as the center of the feature cluster,so as to reduce the intra-class distance of each fault category and enable the model to learn the features with high correlation with the fault categories.Experiments on two datasets verify that the model can improve the cross-domain generalization ability and realize bearing fault diagnosis under varying conditions.In order to further characterize the features information of bearing vibration signals in multiple source domains,a multi-source alignment domain adaptation for bearing fault diagnosis method is proposed for the multi-source domain fault diagnosis problem.For each combination of source domain and target domain,domain adaptation is performed separately.The learned features are input to the corresponding source domain classifier.The classifier discrepancy loss is introduced to minimize the difference between the output of each source domain classifier,so that the classification results of the classifiers in the target domain tend to be consistent.The average output probability of each classifier was calculated as the diagnosis result.Experiments on variable-load and variable-speed bearing data sets prove the validity of the model.By comparing existing single-source domain adaptation and multi-source domain adaptation methods and ablation experiments,it is proved that the proposed method can make full use of fault classification information in multiple source domains.In the case of unknown target domain labels and lack of prior knowledge,the model can bring more balanced data distribution and higher classification accuracy.