Research On Rolling Bearing Fault Diagnosis Based On Deep Transfer Learning

Rotating machinery has been widely used in many important engineering fields,such as aerospace,rail transit,petrochemical,and so on.Rolling bearings are one of the key components of rotating machinery.Rolling bearings usually work in complex and harsh environments,and have characteristics such as heavy loads,high temperatures,corrosion,and high operating rates to meet actual production needs.In addition,complex mechanical structures and constantly changing operating conditions lead to the failure of rolling bearings.Damage to rolling bearings can bring many serious consequences,such as reduced performance and motion accuracy,increased vibration and wear,and even lead to production downtime and casualties.Therefore,it is of great significance to study a more scientific and efficient fault diagnosis method for rolling bearings.The large data of rolling bearings monitored on site for a long time often have few fault samples and are affected by noise and monitoring errors.Useful fault information can easily be submerged in normal sample data,especially for the problem that data sets under varying operating conditions and different scenarios follow different distributions.If the fault diagnosis method based on deep learning is directly used for identification,it is easy to cause erroneous or missed judgments,in order to solve the continuous changes in actual work conditions Due to the scarcity of fault tag data,this paper proposes a deep migration method based on adversarial networks,which is a deep migration bearing fault diagnosis method based on Joint Generalized Sliced Wasserstein Distances.This method reduces the joint distribution distance of labeled data and unlabeled data through domain confrontation,and extracts transferable features in the source domain,Realize unsupervised adaptive intelligent fault diagnosis.The main research content and results of this article are as follows:(1)Aiming at the problem of label data scarcity in field rolling bearings under varying operating conditions,a domain adaptive deep migration fault diagnosis model is proposed.Most existing methods are to reduce the edge distribution of source and target domain features and ignore the calculation of conditional distribution.This method is an improvement on the domain adaptive deep migration learning method,adding a structure of feature joint distribution calculation,This calculation can extract transferable features in the source domain,optimize the domain shift problem caused by different feature distributions in the source and target domains,enhance the generalization ability of the model,and achieve adaptive intelligent fault diagnosis.The bearing failure simulation signal and bearing experimental data from Case Western Reserve University and Xi’an Jiaotong University were used to verify it.The results showed that the migration accuracy of bearing data from Case Western Reserve University and Xi’an Jiaotong University under different working conditions reached 100%,and the migration accuracy of bearing data from different sampling frequencies and bearing models reached 95% on average,The task accuracy rate of migrating bearing data from Case Western Reserve University to Xi’an Jiaotong University’s same frequency bearing data reached 84.6%.(2)A new deep migration bearing fault diagnosis method based on joint generalized slice Wasserstein distance is proposed to solve the problems of missing labels,complex joint distribution calculations,and gradient disappearance in target domain sample data.The algorithm proposes the concepts of generalized slice Wasserstein distance and top K correlated pseudo labels,which enhance the robustness and feature extraction ability of the model.The experimental data of bearings from Case Western Reserve University and Xi’an Jiaotong University were used to validate it.The results showed that the migration accuracy of bearing data from Case Western Reserve University and Xi’an Jiaotong University under different working conditions reached 100%,and the migration accuracy of bearing data from different sampling frequencies and bearing types reached 98% on average.(3)In order to further validate the Wasserstein distance-based deep migration bearing fault diagnosis method based on the joint generalized slice,the application research work based on deep migration learning was conducted using the actual working data of reciprocating compressors.The transfer learning was conducted using the experimental bearing data of Case Western Reserve University and Xi’an Jiaotong University as the source domain,and the actual working data of reciprocating compressors as the target domain.The results show that the average accuracy rate of migrating experimental data to actual work data reaches97.56%.