Intelligent Diagnosis Of Rotating Machinery Based On Domain Adaptation And Domain Generalization

It is extraordinary vital to safeguard the stability and reliability of the operation of rotating machinery,on account of these kind of components such as rolling bearings and gears are extensively applied in industrial equipment.Nowadays,mechanical equipment become further sophisticated,electronic and automatic with each passing day which led to the mechanical structure of related equipment getting incremental precise and complicacy.Hence,the demand for equipment fault prediction,diagnosis and maintenance is becoming more and more urgent.Because of the random and accidental characteristics of mechanical equipment failure,fault data is difficult to obtain.It is difficult to train the intelligent fault diagnosis technology based on big data,and the fault label is difficult to allocate.The lack of data and fault tag has become an important problem restricting the development of fault diagnosis.The cross domain learning technology based on domain adaptation and domain generalization can solve the problem of data shortage,and can make use of a few existing fault data efficiently and give full play to the optimal diagnosis effect.(1)For the sake of finding a sterling feature representation across two domains,a method based on transfer component analysis is proposed to apply to domain adaptation.In reproducing kernel Hilbert space,the maximum mean discrepancy is applied to learn the common transfer components between domains.In the spanned subspace of transfer components,the difference of data distribution between domains is abated.Through the feature representation of subspace,the trained classifier in the source domain could be apply to the target domain.Then a fast representation learning algorithm compatible with domain adaptation and domain generalization is proposed: scatter component analysis.By using scatter to quantify the key indexes,the optimization problem is transformed into solving generalized eigenvalues,so as to obtain a fast and accurate solution and minimize the mismatch between two domains.(2)The classification performance is degraded due to the scarcity of bearing fault data,the distribution difference between the samples and the mismatching of the data.Based on TCA(transfer component analysis),a domain adaptive intelligent fault diagnosis technology for bearings is proposed.The inter domain feature representation is obtained and the cross domain feature fault information migration is completed.The feature distance between different distribution is reduced by minimizing the mean difference in feature subspace.Therefore,the model can achieve relatively high classification accuracy only by a small number of target data.The diagnosis technology is used in the fault data set of rolling bearing under variable working conditions.It can achieve high accuracy and good classification performance without parameter adjustment of the model.This method makes intelligent fault diagnosis technology more practical.(3)In this paper,a unified domain adaptive and domain generalization model,scatter component analysis(SCA),is constructed and applied to the intelligent diagnosis of rolling bearings.As a fast representation learning algorithm,SCA can be used in both domain adaptation and domain generalization.It is based on a geometric measure,namely scatter,and is used in reproducing kernel Hilbert space.Firstly,the algorithm quantizes the scatter and finds a balanced representation between maximizing the divisibility of classes,maximizing the separability of sample data and minimizing the mismatch between domains.Then,the intelligent diagnosis algorithm based on SCA is applied to the bearing vibration signal under variable conditions.It can quickly solve a generalized eigenvalue problem and get an accurate solution,which improves the ability of domain adaptation and domain generalization.The experimental results show that the proposed algorithm can provide advanced rolling bearing fault identification accuracy,effectively solve the problem of sample data deviation,identification efficiency and performance is much faster than previous algorithms.