Research On Intelligent Diagnosis Of Rolling Bearing Fault Based On Sparse Learning

The rapid development of modern industry makes large rotating machinery equipments be applied continuously.As the basic component of rotating machinery,once rolling bearing exhibits faults.If they can not be found and take corresponding measures timely,they will cause significant economic losses and even casualties.Therefore,the research on intelligent diagnosis method of rolling bearing has great engineering practical significance.Due to the repeatability and sparsity of fault feature signal in the whole signal,sparse representation is suitable for solving the problem of mechanical fault diagnosis,and provides a powerful tool for early fault feature extraction and classification based on vibration analysis.The research problems include: the sparsity of sparse coefficient is more considered,but the effectiveness is less considered;dictionary construction is lack of consideration on fault feature and weak adaptability;the selection of key parameter in optimization algorithm is not only fixed,but also depends on empirical value mostly.Based on signal sparse representation,the thesis makes the in-depth study on rolling bearing fault diagnosis around the above problems,which is the key and prone to failure component.The main research contents are as follows:For fault feature extraction of rolling bearing based on sparse representation,more consideration is given to the sparsity of sparse coefficient,but less consideration is given to the effectiveness,the feature extraction method based on Generalized Minimax Concave(GMC)penalty term was proposed.Once the regularization parameter of optimization algorithm is set,it lacks the adaptability to deal with different types of fault signals,and ignoring the distinguishability between them based on sparse reconstruction,thus affecting the accuracy of fault classification.The Relaxation and Compactness factor(RC)adjustment strategy with the ability to adaptively adjust the regularization parameter was designed,which improves the optimization algorithm,and the distinguishability of different types of fault signals is promoted.The fault feature extraction method based on Tunable Q-factor Wavelet Transform(TQWT)was proposed,for the problem that the analysis dictionary constructed by wavelet basis function often lacks the consideration of fault characteristics,which affects the representation ability of sparse coefficient for fault signal.Using fault characteristics to construct wavelet transform basis function not only improves the sparsity,but also enhances the representation ability of sparse coefficient for fault signal.In addition,considering the disadvantage of selecting sparse coefficient based on the energy commonly,by means of the sensitivity of Spectrum Kurtosis(SK)to impulse signal,a series of subband wavelet sparse coefficients were selected to reconstruct fault feature signal with prominent components of fault signals.It improves the rational extraction of sparse coefficients after TQWT,and provides guarantee for better fault classification in the next step.In order to the problem that current fault classification based on sparse representation often directly adopts the training samples to construct a fixed classification dictionary,which is lack of accuracy and adaptability,the fault classification method based on Two-Phase Test Sample Sparse Representation(TPTSR)was proposed.In the framework of sparse representation,the specific classification dictionary is constructed for each sample to be classified based on the nearest neighbor principle,improving the matching ability of classification dictionary to the sample,in the first stage.In the second stage,fault classification is realized based on the current specific classification dictionary.The two-stage progressive sparse representation method not only overcomes the problem of single fixed dictionary for sparse representation classification,but also improves the accuracy of fault classification.Aiming at the problem that the regularization parameter is commonly set based on the empirical value in the process of solving sparse coefficient,the research on setting the parameter value is carried out,based on deep learning training regularization parameter regression network.Taking sparse representation model as the framework,deep learning network is embedded to train the regularized parameter regression network,and then the fault intelligent diagnosis method based on deep sparse representation was proposed.This method combines the model-based sparse representation method and the data-based deep learning method,which realizes the single module fault diagnosis method.