Research On Rolling Bearing Fault Diagnosis And Prognosis Method Based On Deep Learning

Rolling bearing,one of the most common types of bearings,often runs in the complex environment of high temperature and high load.It is inevitable that bearing break down from time to time.Due to mechanical rolling bearing main effect reduces the internal friction loss,the rolling bearing failure has a great impact on the operation of machinery,even a serious safety accident is happened,leadings to loss of economy and life property.The technical research of fault diagnosis and prognosis is needed to ensure the safe operation of machinery.Among many research methods,the data-driven diagnosis and prediction method has achieved remarkable results in recent years.Such a diagnosis method starts to study the occurrence of rolling bearing faults from the perspective of data analysis,and such a prediction method can more easily explore the variation law of bearing performance after the occurrence of faults.However,timefrequency domain analysis and others methods are difficult to extract feature to varying degrees now.This thesis aims to use deep learning technology to further improve the recognition accuracy of fault diagnosis of rolling bearing and the fitting precision of fault prediction,fully explore the deep learning new innovation for the two pieces of research content,the main content of the paper is as follows:First of all,the hyperparameter optimization algorithm is studied.A NLF optimization algorithm is proposed to balance local development and global exploration in optimization based on random search algorithm.Through the comparative analysis with random search algorithm,demonstrate the NLF algorithm feasibility of the proposed for the next of the fault diagnosis based on LSTM research to provide reliable basis.On this basis,a fault diagnosis method based on NLF-LSTM is proposed,and the results of CWRU data set show that the accuracy of fault classification and recognition of this method is further improve.Secondly,the principle of S-time entropy is studied and applied.When the coefficient matrix obtained after S transformation is used in the calculation of S-time entropy,de-valuing operation is carry out for each column first to improve the calculation speed.Based on this,a fault prognosis method based on LSTM is proposed.In order to verify the effectiveness of the method,the IMS data set is introduced to carry out experiments.Through the comparison of various machine learning methods,it is shown that the proposed method significantly improves the prediction accuracy and has a good fitting effect on the performance degradation trend.