Transfer Learning-based Methodologies For Fault Diagnosis And Remaining Useful Life Prediction Of Rotary Machine

At present,all kinds of machinery and equipment used in industrial production,aerospace and rail transit are becoming more and more complex,and the level of automation and informatization is getting higher and higher.Once the gears,bearings and other mechanical rotating parts are damaged,it will lead to mechanical failure,production stagnation and even human casualties.Therefore,it is of great significance to establish an effective diagnosis and prediction model for rotating machinery system.As the mechanical fault diagnosis and useful life prediction technologies based on the new artificial intelligence become more and more mature,more complex and precise mechanical damage model can be built,so as to improve the performance of fault diagnosis and remaining useful life prediction.However,there are some defects in the existing models,such as high algorithm complexity,strict requirement of modeling data and low model generality.To overcome the shortcomings of the existing models,this paper takes rotating machinery as the research object and transfer learning as the theoretical basis to carry out intelligent fault diagnosis and remaining useful life prediction,as follows:(1)The application of transfer components,transfer domains and transfer times during transfer learning are described in the mechanical vibration signals.During transfer components,the sample transfer model and the feature transfer model are compared with each other.During transfer domains,the cross channels transfer model and cross devices transfer model are compared with each other.During the number of transfers,the direct transfer model and the transitive transfer model are compared with each other.Meanwhile,the basic application conditions of each model will be analyzed.(2)From the perspective of different transfer components of mechanical fault diagnosis,the sample transfer model based on probability factor Tr Adaboost algorithm and the feature transfer model based on local vector transfer are proposed,respectively.Three main steps are involved in the sample transfer model,including the singular value decomposition-based vibration feature extraction,the sign rank/Chi-square test-based serial domain similarity index,as well as the probability factor Tr Adaboost algorithm.Two main steps are involved in the feature transfer model,including the multi-domain feature extraction and local vector transfer.The experimental results show that the sample transfer model is superior to the feature transfer model in fault location diagnosis,while the latter is superior to the former infault diameter diagnosis.Under the premise of ensuring the diagnostic accuracy,the iteration speed of the probability factor Tr Adaboost algorithm is effectively improved compared with that of Tr Adaboost algorithm.Meanwhile,the local vector transfer algorithm is helpful to solve the problem of uneven quality of multi-domain features of mechanical vibration signals.On the whole,the fault diagnosis accuracy of sample transfer model is higher than that of feature transfer model.(3)From the perspective of different transfer tasks of mechanical fault diagnosis,the cross channels transfer model based on penalty domain selection machine(PDSM)and the cross devices transfer model based on weak constraint fast self-organizing mapping(FSOM)are proposed,respectively.Two main steps are involved in the cross channels transfer model,including band-selective independent component analysis(BS-ICA)as well as PDSM domain adaptation algorithm.Four main steps are involved in the cross devices transfer model,including the real-imaginary polar diagrams,the weak constraint FSOM,the channel selection of source device based on cross Minkowski distance and the channel fusion of target device based on second-order selective ensemble.The experimental results show that the simple mechanical parts are more likely to be transferred to complex mechanical parts in both cross channels or cross devices transfer models.Meanwhile,the difference between healthy gears is the least,and the difference between light cracked gears is the greatest when transfer between two devices.Compared with the DSM model,the domain penalty factor vector and the signal penalty factor vector are introduced in the PDSM model to improve the fault diagnosis accuracy and real-time performance in the cross channels transfer model.Compared with the FSOM algorithm,the weak constraint FSOM algorithm weakens the constraint conditions of the initial map and enhances the diagnostic adaptability of the cross devices transfer model.On the whole,the fault diagnosis accuracy of the cross channels transfer model is higher than that of the cross devices transfer model.(4)From the perspective of different transfer styles of remaining useful life(RUL)prediction,the direct transfer model based on transfer compact coding for hyper plane classifiers(TCCHC),as well as the transitive transfer model based on the domain independent support vector machine(DI-SVM)algorithm is proposed,respectively.Three main steps are involved in the direct transfer model,including the mechanical fatigue degradation curve based on mel-frequency cepstral coefficient(MFCC),the TCCHC curve transfer algorithm,as well as exponential semi-random semi-random extended Kalman filter(EKF).Four main steps are involved in the transitive transfer model,including mechanical fatigue degradation curve based on the joint evaluation index,the early fault detection based on maximum correlation kurtosis de-convolution(MCKD),the domain selection based on the comprehensive domain discrimination index,and the middle-domain DI-SVM algorithm.The experimental results show that the direct transfer model is better than the transitive transfer model when the distance between the source domain and the target domain is small and the middle domain is far.On the contrary,the latter is better than the former.Meanwhile,the TCCHC algorithm helps to reduce the difference between the fatigue degradation curves of the source and target devices.Compared with single DI-SVM algorithm,the life prediction error when the middle domain is introduced is smaller,but the latter depends on the quality of the middle domain.On the whole,the RUL prediction accuracy of transitive transfer model is higher than that of the direct transfer model.