Research On Fault Diagnosis Of Rolling Bearing Based On The Deep Auto-Encoder Network

Rolling bearing is one of the most important parts in mechanical equipment.Its health status directly relates to the equipment’s continuous and reliable operation.Once the rolling bearing breaks down,it will cause economic losses or even a disaster.Therefore,the research of rolling bearing fault diagnosis technology is of great significance.The fault diagnosis process is generally divided into four steps:fault data collection,feature extraction,feature selection and fault identification.Fault feature extraction is a key step in the diagnosis process.This step is generally solved by signal processing methods,which strongly depend on professional knowledge and technical personnel.The rise of Artificial Intelligence techniques has opened up a new world for rolling bearing fault diagnosis.Deep learning is an emerging force in the field of artificial intelligence.It can automatically learn the intrinsic characteristics from the data to obtain a good feature expression,which provides new ideas for fault feature extraction and diagnosis of rolling bearing.This paper focuses on the application of a deep auto-encoder network in the fault diagnosis of rolling bearings.The main research contents are as follows:Based on the detailed analysis of the principle of deep auto-encoder network,its performance of feature extraction and classification ability is studied from data reconstruction and classification rate.DAE reconstruction network is set up and the performance of feature extraction is initially verified from the perspective of data reconstruction.An external study of comparison in signal classification with traditional feature extraction methods that are PCA and KPCA validates the superiority of DAE’s performance.DAE diagnosis model is constructed.As to the decision of the key parameters,The simulated data was used to study the number of network layers and nodes of hidden layers.The optimal structural parameters were determined from the classification rate and computational cost.After that,the diagnosis model was applied to the public rolling bearing data set under multiple working conditions,achieving ideal performance.Meanwhile,as to the problem that normal samples usually larger than fault samples,the performance of the model under unbalanced samples was also explored.SVM and BP network model are also compared with DAE model,which indicates that the constructed DAE model could identify different fault severities and fault orientations with more stability.Given the problem of insufficient target data samples in actual situations,this paper proposed an enhanced model,which introduces transfer learning into DAE networks,using the public data and collected data as assistance and target respectively.The result shows that the performance of the improved model is better than the standard model when there are few target data samples.