Fault Diagnosis Method Of Rotating Machinery Based On Unsupervised Transfer Learning

In industrial practice,the working environment of rotating machinery(such as engine,steam turbine,etc.)is harsh and changeable,and its mechanical properties will slowly decline.If the corresponding measures are not taken in time in the stage of performance degradation,the equipment will be shut down,seriously,will cause safety accidents.Therefore,it is of great practical requirements and practical significance for key parts in rotating machinery to carry out fault diagnosis.Because of the emergence of artificial intelligence technology,deep learning theory is applied by many scholars to the field of mechanical fault diagnosis,which is purposed to automatically learn fault feature information from a large number of historical data.These fault diagnosis models not only have high fault diagnosis accuracy,but also are very intelligent.However,in practice,because the working condition of rotating machinery often changes,it is difficult to directly monitor the state data under the current actual working condition,which leads to obtain samples of known fault types under the current working condition difficultly;In addition,due to the transient feature of working conditions,the distribution characteristics of historical fault samples(i.e.,the labeled samples in auxiliary domain)under previous working conditions are obviously different from those of the samples in target domain.Because both traditional machine learning and deep learning are based on the assumption that training samples and testing samples follow the same or similar distribution,the generalization ability of fault diagnosis model based on traditional machine learning and deep learning is poor,so it is not suitable for fault diagnosis under variable working conditions.The transfer learning theory developed in recent years provides a new solution for fault diagnosis of rotating machinery under variable working conditions.Based on the theory of transfer learning,in view of the shortcomings of rotating machinery fault diagnosis methods based on traditional machine learning and deep learning,in this paper,manifold subspace transfer learning and depth transfer learning are deeply researched to improving the fault diagnosis accuracy and time calculation efficiency of rotating machinery fault diagnosis method.(1)Aiming at the large difference between the distribution of training samples and the testing samples under variable working conditions,which leads to the problem of low fault diagnosis accuracy of rotating machinery fault diagnosis methods based on traditional machine learning,a new unsupervised transfer learning method ——Manifolds subspaces transfer learning(MSTL)is proposed for rotating machinery fault diagnosis.In MSTL,the uniform alignment cost function of subspace manifolds is constructed to align the subspaces and low-dimensional manifolds of the two domain samples at the same time.This can not only make full use of the category information of the auxiliary domain samples and mine the variance information of the target domain samples,but also make the differenc of subspace and data structure is the smallest after the samples in two domains are mapped,which enhances the domain adaptability of MSTL;the strategy of using the minimum the joint distribution subspace manifold unified alignment total loss function values for feature transfer can improve the accuracy of fault classification after transferring.The above advantages of MSTL make it possible to use the labeled samples under historical operating conditions to perform high-precision fault diagnosis on the current testing samples of rotating machinery when the labeled samples under the current operating conditions do not exist.The fault diagnosis example of rolling bearing verifies that proposed method has higher fault diagnosis accuracy and computational efficiency than the two current mainstream transfer learning methods.(2)In view of the problem that the transfer learning method based on subspace modeling needs to rely on manual extraction of fault features,a new unsupervised transfer learning method——deep convolution domain against transfer learning(DCDATL)fault diagnosis method.In the proposed DCDATL,a new deep convolution residual feature extractor is constructed to extract high-level features,which can avoid gradient problems such as gradient disappearance and gradient divergence during training DCDATL,thus improving the convergence and non-linear approximation ability of DCDATL.At the same time,the joint distribution of labeled samples in auxiliary domain and unlabeled samples in target domain is creatively used for domain-adversarial training,which can enhance the adaptability of samples in auxiliary domain to target domain and improve the transfer performance of DCDATL.Moreover,the strategy based on minimizing the joint distribution domain-adversarial total loss function of DCDATL is innovatively presented to improve the fault classification accuracy after high-level feature transfer.The above advantages of DCDATL make it feasible to perform high-precision fault diagnosis on current testing samples by using the historical labeled samples in auxiliary domain when there are no labeled samples in target domain in the case of not need to rely on manual extraction of fault features.The rolling bearing fault diagnosis example verifies that this method has higher fault diagnosis accuracy and computational efficiency than the MSTL proposed above and the other two popular transfer learning methods.(3)In view of the above-mentioned transfer learning method based on the deep adversarial network,the feature learning ability will decrease when the samples in auxiliary domain and target domain are interfered by large noises.A new unsupervised transfer learning method is proposed — — Deep convolution domain-adversarial manifold transfer learning(DCDAMTL)for fault diagnosis of rotating machinery.In DCDAMTL,covariance pooling operation is used to make the high-order statistical information of the signal not lack,so that DCDAMTL can better capture the subtle regional features of the signal.The symmetric positive definite manifold network is designed for feature extraction,so that the geometric information of the data structure of the two domain samples is not lost,and the difference of the data structure is minimal,which improves the feature learning ability of DCDAMTL.Compared with proposed MSTL,DCDATL and other two popular transfer learning methods,DCDAMTL has higher fault diagnosis accuracy and better computational efficiency in the case of large noise interference in auxiliary domain and target domain samples.(4)Based on the above theoretical research results,a rotating machinery fault diagnosis software system based on unsupervised transfer learning is developed by using Python development language and My SQL database development platform.The software system includes automatic extraction of original data,selection of diagnosis method,parameter setting,output and storage of results,etc.It has the characteristics of fast diagnosis,complete core functions,friendly interface and strong operability,providing good software technical support for fault diagnosis of rotating machinery.