Deep Fault Diagnosis Network For Transmission Rotating Parts Under Complicated Working Conditions

Mechanical big data has the characteristics of low information density,large capacity,diversity,high speed and multi-source heterogeneity.Traditional vehicle fault diagnosis methods rely on a lot of expert experience,which cannot meet the demand for automatic and efficient feature extraction.Meanwhile,the working condition variation and dataset class imbalance of vehicle equipment also need to be solved urgently.In order to meet the diagnostic requirements of mechanical transmission rotating parts,this paper conducts in-depth research on intelligent fault diagnosis methods for transmission rotating parts(gears and bearings)based on deep learning and transfer learning theories.The main contributions of this paper can be summarized as follows:(1)Aiming at realizing fast and accurate fault diagnosis of vehicle transmission rotating parts,an adaptive overlapping convolutional neural network(AOCNN)framework is constructed.AOCNN can not only extract shift-invariant features from raw signal samples automatically,but also accelerate the training process through a parallel convolution layer.Meanwhile,the network with AOCNN framework only needs to train its local feature extraction network,which reduces the sample demand in network training,and further improves the training speed and generalization ability of the network.The experiments of gear and bearing fault diagnosis validate that the AOCNN with Sparse filtering or Autoencoders as local feature extraction part can utilize a very small number of raw signal samples to obtain a highly precise fault diagnosis network quickly.(2)Aiming at the hot fault-sensitive sparse feature extraction problem,parameterized sparse label matrix(PSLM)is designed to introduce feature sparsity through the supervision mechanism.Further,based on PSLM and topology structure of the feature matrix,supervised sparse filtering(SSF)is proposed to extract highly fault-sensitive sparse features.SSF converges quickly and the fault-sensitive features extracted by SSF can also improve the fault diagnosis accuracy and stability.The high efficiency of SSF in fault-sensitive feature extraction is verified by specially-designed bearing fault datasets containing samples from fifteen categories of health conditions.Investigations also show that SSF can be fused with AOCNN easily to boost the diagnois performance further utilizing raw signal samples.In addition,PSLM can also introduce sparsity into feature extraction networks such as Autoencoders and enhance their sparse feature extraction capabilities.(3)As common intelligent fault diagnosis methods of vehicle transmission rotating parts fail in certain degree when meeting class imbalance problems,we investigate the degrading problem of sparse filtering(SF)when meeting class imbalance,and propose balanced sparse filtering(BSF).From the perspective of cost sensitivity analysis,BSF first alleviates the network parameter updating imbalance via the proposed balancing matrix,and then improves the feature distinguishability of rare classes by minimizing the feature sparsity divergence among different classes.Vast case investigation on a bearing dataset validates the effectiveness of the proposed method in restraining the performance degradation caused by class imbalance problems.(4)Aiming at the fault diagnosis of vehicle equipment with frequent working condition variation,we firstly introduce a transfer learning method called adaptive batch normalization(Ada BN),then combine it with a deep network called stacked autoencoders(SAE)and construct Ada BN-SAE.Ada BN-SAE enhances the sharing features of the same class under different working conditions by aligning the marginal distributions of features in each feature layer,which improves the network robustness to working condition variation.In the specially-designed gearbox fault diagnosis,Ada BN-SAE greatly improves the diagnosis accuracy and stability,and expands the applicability of network in practical applications.At the same time,through the investigation of features in different layers,it is found that both the sharing and distinguishing properties of the features rise with the increase of network depth in Ada BN-SAE.(5)It is often the situation that complex working condition variation of vehicle transmission rotating parts will lead to complex joint distribution shift in new datasets.Targeting at the problem that existing methods for fault diagnosis under working condition variation only focus on marginal distribution aligning,adaptive joint distribution alignment(AJDA)method is proposed,and based on AJDA,a deep transfer network(DTN)is also constructed.AJDA realizes more smooth conditional distribution aligning through soft labels instead of existing pseudo labels.Meanwhile,it introduces dynamic distribution adaptation(DDA)to realize dynamic adaptation for variable joint distribution caused by different working conditions.It is also found in experiments that the features extracted via the DTN are highly sharing and sparse,which enables more precise and stable fault diagnosis under complicated variable working conditions,and further improves the robustness of the network.(6)Improved joint distribution alignment(IJDA)method is put forward to explore a more practical vehicle fault diagnosis problem,where working condition variation and class imbalance appear simultaneously.IJDA first combines data augmentation and Gaussian white noise in generating pseudo samples to solve class imbalance problem.Then,it constructs a two-stage feature extraction network with AJDA to extract sharing fault-sensitive features from datasets with a large range of rotation speed oscillation quickly.Finally,it realizes robust fault diagnosis with Softmax regression trained by a dataset from only one working condition.In experimental validation,the specially-designed bearing datasets with a large range of rotation speed oscillation are used.It presents that IJDA trained by the datasets from only two working conditions can obtain precise and stable fault diagnosis results on new datasets within a large range of new working conditions,which realizes more practical and accurate fault diagnosis.