Application Research On Fault Diagnosis Of Rolling Bearing Based On Deep Auto-encoder Network

As a key transmission component of rotating machinery,rolling bearings play an important role in determining the operating state of the equipment.However,under complex working conditions,rolling bearings are prone to wear,pitting,adhesion and other failures.If failures cannot be found in time during mechanical operation,serious losses will occur.In view of this,online monitoring of rolling bearing status and early detection of faults are of great research significance.Traditional machine learning-based fault diagnosis methods for rolling bearings rely too much on expert experience to manually extract features,which is time-consuming and labor-intensive.Deep learning can use the deep structure to automatically learn high-level essential features and establish a complex non-linear mapping relationship between input and output.AS one of the more commonly used models in deep learning,regardless of whether the input is preprocessed,deep auto-encoder network only executes it as a low-level feature.And then it automatically extracts more discriminative features based on the input.Aiming at the problems of insufficient feature extraction ability,difficulty in parameter selection and poor generalization capabilities of traditional deep auto-encoder networks,this paper optimizes its parameters and structure to improve its feature mining capabilities and then apply them to rolling bearing fault diagnosis.The main research contents of the thesis are as follows:(1)Aiming at the problems of long training time and unstable performance of some existing intelligent diagnosis algorithms,a new method based on ensemble empirical mode decomposition(EEMD)and deep sparse auto-encoder network(DSAE)is proposed.First,the bearing signal is decomposed by EEMD to obtain the Intrinsic mode function(IMF)component.Secondly,the effective IMF component is selected as the feature domain based on the kurtosis value,then the time domain and frequency domain features are extracted to construct a new data set as the input of the diagnosis network.Finally,the new data set is used as the input of DSAE for training and testing.By comparing with traditional intelligent diagnosis algorithms-Support Vector Machine(SVM),Back Propagation Neural Network(BPNN)and DSAE,the results show that the proposed method has superior performance in terms of accuracy and computational time-consuming.(2)To solve the problems of difficult selection of key parameters and insufficient feature extraction capabilities in deep auto-encoder networks,a new method based on cuckoo search optimized deep auto-encoder network is proposed for fault diagnosis of rolling bearings.First,sparse penalty term and contractive penalty term simultaneously add to the loss function of auto-encoder to design a new deep auto-encoder network-Deep Contractive Sparse Auto-Encoder(DCSAE).At the same time,batch normalization is added to the hidden layer to train the network weights more better.Secondly,the cuckoo search algorithm is used to adaptively select the optimal key parameters of the network.Finally,the proposed method is applied to the fault diagnosis of rolling bearing vibration signal.The experiment results of the two data sets show that the proposed network can not only effectively extract fault features,but also has higher classification accuracy.(3)To enhance the ability of the deep auto-encoder network to mine features from the original vibration data,an ’end-to-end’ intelligent fault diagnosis method based on the deep auto-encoder network is proposed.It directly uses the original time domain signal of the bearing as input to achieve integration fault diagnosis of rolling bearings.First,the maximum correlation entropy is used as the loss function of the auto-encoder.Besides,a non-negative constraint is introduced on the network weight to further reduce the reconstruction error.Second,dropout added to the hidden layer of the network to prevent overfitting and reduce computing time.Then,the key parameters of the network are selected adaptively through the gray wolf algorithm.Finally,the proposed method is applied to the measured data.Experiments with two different datasets verify that the proposed intelligent fault diagnosis method has higher robustness and generalization ability to unprocessed raw data samples.