Research On Machinery Fault Diagnosis Method Based On Deep Auto-encoder

Fault diagnosis is important for improving the reliability of machinery,ensuring the long-term stable operation of equipment,and reducing the economic loss caused by downtime.Traditional fault diagnosis methods require experts to extract features manually.However,due to the signal characteristics of transient,nonlinear and non-gaussian,it needs a large amount of analysis and comparisons to get the effective fault features,which is a time-consuming and labor-intensive process.Therefore,how to efficiently extract the discriminative fault features is challenging and becomes a hot topic.In order to extract fault features automatically,this paper devotes to introducing deep auto-encoder(DAE),namely a deep learning method,into the fault diagnosis field.Facing the enough fault samples,we propose a DAE based fault diagnosis method.Next,considering two influence factor,namely signal noise and small sample,we improve DAE correspondingly and then propose the fault diagnosis method.Based on the above two factors,the fault diagnosis methods for noisy signals under small samples are studied respectively.The main contents are as follows:Facing enough fault samples,this paper proposes a sparse DAE(SDAE)fault diagnosis method.For the problem that SDAE will extract similar features,we design a Subset method whose role is to help SDAE to extract the discriminative features from different fault patterns.Additionally,in order to get the optimal parameter configurations,we propose a parameter optimization framework based on particle swarm optimization algorithm.The proposed method is verified on the fault diagnosis of motor bearing(simulation experimental platform)and self-priming centrifugal pump(real machine equipment).Experimental results show that the proposed method can effectively extract the fault feature.Comparisons with other methods show that the proposal has better diagnostic accuracy.Considering the problem that it is difficult to extract effective fault feature from noisy signals,this paper proposes an ensemble DAE based fault diagnosis method.For the feature extraction,we design a CSDAE(Contractive Sparse DAE)+ FDA(Fisher Discriminant Analysis)model to obtain the maximumly detachable features from noisy signals.Aimint at the problem that single model is unreliable,fifteen CSDAE models with different characteristics are designed.On this basis,a weighted voting method is designed to integrate the diagnosis results of multiple models for improving diagnosis accuracy.The proposed method is verified on the fault diagnosis of the self-priming centrifugal pump and motor bearing.Some variants of DAE and the 15 single CSDAE are compared with the proposal.Experimental results show that the diagnosis accuracy of the ECSDAE is significantly better than the compared methods when dealing with noisy signals.Considering the problem that fault samples are small samples and insufficient fault information will lead to low diagnostic accuracy of the fault sample,this paper proposes a fault diagnosis method based on oversampling method and SDAE.For the problem that the existing oversampling methods generate wrong or reduandant samples,a new weighted minority oversampling method(WMO)is proposed for data enhancement.Considering the great influence of outlier for the SDAE performance,the maximum cross entropy criterion and relative entropy are adopted to modify the cost function of SDAE to improve its robustness of dealing with abnormal data.We verify the proposed method on 25 University of California Irvine(UCI)benchmark dataset and compare it with 5 famous small sample learning methods.Besides,two engineering applications including NASA bearing(simulation experimental platform)and wind turbine blades(real machine equipment)are used to compare the proposed method with other 10 methods.The results show that the proposed method is superior to the comparison methods.Additionally,experiments are carried out by setting different amounts of small samples.The results show that the fewer of the small samples will cause the less sensitivity to fault samples and it will give more false alarms.All the results indirectly indicate that the study on fault diagnosis under small samples is of great significance.Considering the above two influence factors,we propose a diagnosis method by combining WMO with Ensemble DAE.For the situation that small samples are from more than one class,we extend WMO using “one-against-all” strategy to realize the data enhancement.Next,we integrate WMO with ECSDAE and propose a new traing method.In addition,the upper and lower error bound of the ensemble method are theoretically analyzed,which proves that the ensemble method is more reliable and effective.Likewise,we test the proposal on NASA bearing wind turbine blades.The results show that the proposed method can identify more fault samples,and the overall reliability of the diagnosis results is higher than the comparison mehtods.Finally,the summarization and the future research are presented.