Research On Bearing Fault Diagnosis Based On Stacked Autoencoder

The rolling bearing is an important part of the mechanical equipment which supports the rotating parts.The health condition of rolling bearings determines whether machinery and equipment can run safely and stably.Therefore,it is of great significance to carry out fault diagnosis research on rolling bearings.In the process of fault diagnosis of rolling bearings,the effective extraction of features is the critical step.Whether the extracted features are representative will directly affect the final fault diagnosis results.Traditional feature extraction methods based on signal processing and machine learning shallow model require a lot of professional knowledge and some expert experience,so it is difficult to realize intelligent bearing fault diagnosis.With the rise of deep learning,deep neural network models can automatically extract useful features and directly classify them,which provides new ideas for intelligent fault diagnosis of rolling bearings.This paper focuses on the application of autoencoder under the framework of deep learning in bearing fault diagnosis.The main work is as follows:(1)Aiming at the problem that the features extracted by the traditional stacked autoencoder are not discriminatively different and the traditional stacked autoencoder has low accuracy in the diagnosis of similar faults in bearings,a stacked autoencoder with dynamic feature enhanced factor is proposed to realize the similar fault diagnosis of bearings.Firstly,features are enhanced by introducing competition and enhanced constraints between hidden neurons in the training process of traditional stacked autoencoder.Then,according to the diversity between different fault features and the information amount carried by the encoding features from the input samples,a dynamic feature enhanced factor is designed,which can adaptively extract the features with large differences.Finally,a stacked autoencoder with dynamic feature enhanced factor is constructed by stacking multiple feature enhanced autoencoders,which improves the diagnosis accuracy of bearing similar faults.(2)Aiming at the limitation that the traditional autoencoder is only suitable for bearing vibration signals with the same distribution,a transfer learning method is combined with a stacked autoencoder to propose a fault diagnosis method of bearing under variable working conditions based on domain adversarial stacked autoencoder.In the process of bearing fault diagnosis under variable working conditions,the source domain data is first used to carry out unsupervised pretraining on the domain adversarial stacked auto-encoding network.Then,the source domain data and target domain data are taken as input data,and the domain-invariant features are extracted by using MMD domain distribution difference measurement method and optimized domain classifier confounding loss method to realize fine-tuning of the network.Finally,the trained network model is used to diagnose the bearing fault in the target domain.