Research On Bearing Fault Diagnosis Method Based On Tight Frame Learning And Convolution Neural Network

Rolling bearing is an indispensable transmission component of rotating equipment.It plays a major role in modern industrial manufacture and is extensive used in multiple domains for instance mining,metallurgy and medical treatment.It’s safe,reliable and smooth operation has an extremely important impact on ensuring production efficiency.Compared with other mechanical parts,rolling bearings are naturally superior in friction loss,assembly,working efficiency and other aspects,so they are also very excellent in rotating machinery.But if the bearing fails,it will cause a chain reaction,such as affecting production,causing property loss,or even endangering life.Therefore,the accurate monitoring and diagnosis of the operating status of rolling bearings ahead of time will be vastly assist to the reliable and safe operation of the manufacturing system.This thesis studies the vibration signals collected by sensors of rolling bearings under different working conditions and loads.Starting from different fault types and various fault depths,diagnoses the inner ring,outer ring and rolling body of rolling bearings.The emphasis is on vibration signal feature extraction and classification.In this thesis,two fault diagnosis means are mentioned,experimental results show that the diagnostic accuracy of the proposed method is close to 100%.The major work details are as follows:(1)Firstly,the background and significance of rolling bearing fault diagnosis are introduced,and then the advanced diagnosis technologies at home and abroad are summarized from the two aspects of signal feature extraction and signal classification and recognition in the process of fault diagnosis.Finally,the principle,advantages and disadvantages of fault diagnosis methods are explained.(2)Aiming at the problem that traditional multi-scale transform features such as wavelet transform cannot adaptively solve the fault diagnosis problem,a sparse dictionary learning model under the constraints of tight frame conditions is proposed.The optimization equations for constructing tight frames are solved iteratively through rolling bearing data samples,which are adapted to data decomposition and complete reconstruction of the tight frame.The frequency distribution of the fault signal in each subspace of the tight frame containing the fault mode is different.By finding the energy distribution of the signal in each subspace,the characteristics corresponding to the fault signal can be established.Since support vector machines are greatly affected by model parameters when classifying feature vectors,genetic algorithms are used to optimize the parameters of support vector machines for fault diagnosis.(3)For settle the matter in the course of signal feature extraction and pattern classification in fault diagnosis,the convolutional neural network in deep learning model is used for automatic intelligent diagnosis of bearing faults.Because the convolutional neural network to behave better on two-dimensional image,so synchronous compression is presented in this thesis wavelet transform for time-frequency analysis,signal synchronization is compressed wavelet transform of wavelet transform time-frequency summation of the scale of the figure under the same frequency,at the same frequency coefficient of compression at around,finally through the special mapping time-frequency,time-scale can be converted to obtain more sparse 2-d time-frequency distribution,by building the convolutional neural network model,frequency pair figure for automatic classification,achieve the intelligent fault diagnosis of bearing.