Rolling Bearings Fault Feature Analysis And Diagnosis Method Based On Data-driven Model

As the key component of rotating machinery and equipment,the rolling bearing failure will seriously affect the safe and stable operation of machinery and equipment.The time-frequency analysis of vibration signals is an effective method to realize the fault feature analysis of rolling bearings.However,the redundant and interference information exists in the initial feature set obtained through time-frequency analysis,and the fault state-sensitive features need to be selected.The high dimensionality of the signal feature space requires the use of a research dimensionality reduction mapping method to enhance the state representation of low-dimensional feature space.The traditional pattern recognition method is difficult to adapt to the complex mapping between high-dimensional feature space and state space.The deep learning method has the ability of high-dimensional feature adaptive analysis,which is suitable for intelligent analysis of high dimensional feature space in fault state and related application research in its infancy.In view of the above problems,this paper carries out the following research work.In the aspect of signal analysis and feature extraction,the time-frequency analysis of vibration signal is an effective method to realize the fault diagnosis of rolling bearing.Because of partial redundancy and interference information in the initial feature of time-frequency analysis,it is necessary to select the sensitive characteristic of fault state.In the aspect of dimensionality reduction,after the preliminary feature selection,the fault state feature space dimension is higher.It is necessary to study the spatial mapping method to improve the state recognition ability of the low-dimensional feature space obtained by the feature reduction method.But the traditional model is difficult to characterize the complex mapping relationship between the high-dimensional feature space and the state space of the measured signal,and faced with dimensionality of disaster.As a new method of machine learning,deep learning can automatically extract the required features from a large number of data,and be suitable for diversity,non-linear,high-dimensional data analysis based on the large data.Some related application researches are at the start stage.In view of the above problems,this paper focuses on the following aspects:(1)In this paper,we focus on the time-frequency analysis method of vibration signal based on ensemble empirical mode decomposition(EEMD).EEMD is used to decompose the vibration signal into intrinsic mode function(IMF),and obtain the Hilbert marginal spectrum(HMS)of vibration signal,which contains the frequency distribution of the vibration signal at different scales.We propose to use HMS as the initial feature of the fault diagnosis model.In order to reduce the redundancy and interference information in HMS and reduce the input space dimension,a HMS fault state characteristic frequency band extraction method called WMSC(Window Marginal Spectrum Clustering)based on sliding window and Rand index is proposed to complete the sensitive fault characteristic frequency band in HMS The selection.In this paper,the fault diagnosis model of HMS-WMSC-SVM based on SVM is established.The vibration signal analysis results of the bearing failure experimental bench show that the model can obtain the high accuracy of bearing fault state recognition and has a strong ability of noise suppression.(2)In this chapter,we further study the fault pattern recognition method based on the statistical parameters of IMFs,Hilbert envelope(HES)and HMS obtained by EEMD analysis of vibration signals.Different statistical parameters have different sensitivity to the bearing fault,so it is a key step of constructing fault diagnosis model to select the sensitive feature,which can reflect fault of equipment.In order to solve this problem,we propose a feature selection method called FSASR,which use the ratio of adjusted rand index(ARI)and standard deviation(STD)as the sensitivity index of statistical parameters to realize the quantitative analysis of sensitivity.Support margin-local fisher discriminant analysis(SM-LFDA)based on local fisher discriminant analysis(LFDA)is proposed to solve data redundancy problem in high dimensional characteristic space,which can map statistical feature from high dimensional to low dimensional and improve the accuracy of pattern recognition.Fault diagnosis model,based on support vector machine(SVM),K-nearest neighbor(KNN),ensemble method(EM),is used to analyze the vibration signal of bearing fault,the results show that the FSASR can effectively select the sensitive feature,and the SM-LFDA can improve the accuracy of the fault diagnosis model based on EEMD.(3)The maximum overlap discrete wavelet package transform(MODWPT)is studied,which is a time-frequency analysis method of the vibration signal,the transform coefficients of it are translational invariance,it`s time resolution same on each decomposition layer and it can avoid the phase distort.Single reconstructed signal containing frequency information obtained by rebuilding the wavelet packet nodes decomposed by MODWPT,calculating the statistical parameters of the single reconstructed signal and its HES to construct feature space and the FSASR is used to select sensitive features.Supervised neighborhood preserving embedding with label information(SNPEL)is proposed,which based on a non-supervised popular learning named Neighborhood preserving embedding.Combining SNPEL and pattern recognition method,SVM,KNN and EM,establish the fault diagnosis model.Further,the experimental results show that the FSASR method can effectively select the fault sensitive features,and the SNPEL method can improve the accuracy of pattern recognition model of based on MODWPT.(4)Summarizing the experimental analysis of multiple dimensionality reduction methods and pattern recognition methods can conclude that the dimension reduction method can obtain the low dimension representation of feature space,the pattern recognition method can identify the fault,but different dimension reduction method and pattern recognition method have different analysis ability for feature space.Deep learning has advantages in high dimensional data processing and non-linear data analysis,the fault pattern recognition method used Deep Belief Networks(DBN)is studied.Based on vibration signal analysis,statistical parameter analysis and fault feature selection method,we establish pattern recognition model such as EEMD-DBN and MODWPT-DBN.The experimental results show that DBN can replace the traditional feature dimension reduction and pattern recognition method,and fault diagnosis model based on DBN can effectively improve the accuracy of fault pattern recognition,and have the strong ability of adaptation to FSASR.In summary,the WMSC and FSASR methods are proposed in this paper to reduce the interference and redundant information in the initial eigenspace after time-frequency analysis.SM-LDA and SNPEL methods are proposed to improve the state representation in low-dimensional mapping space.In this paper,we establish a fault condition identification model based on DBN to improve the adaptability of pattern recognition method to feature space and form a complete system of fault analysis and diagnosis based on data-driven rolling bearing.Experimental results show that the proposed method can effectively improve the accuracy of fault identification and has strong adaptability.