Research On Rolling Bearing Fault Diagnosis Based On Preferred Feature

In recent years,with the rapid development of signal processing and artificial intelligence technology,data-driven fault diagnosis method has gradually become a research hotspot.The signal processing and feature extraction in the process of fault diagnosis,as the premise and foundation of fault conditions recognition,will have an important impact on the fault diagnosis results.At present,researchers widely use time-frequency analysis method to process the rolling bearing vibration signals,which is the first step of fault diagnosis.Due to that the rolling bearing vibration signal is nonlinear and non-stationary,after the process of vibration signal by time-frequency analysis method,the dimension of the original feature set is high,and there are problems with interference and redundancy features.In addition,most of the data-driven fault diagnosis models constructed by using traditional machine learning methods have two main problems:(1)the lack of a great quantity of labeled fault training samples;(2)the construction of fault diagnosis model is under the assumption that the training dataset and testing dataset have the same distribution.Therefore,it is difficult for the traditional fault diagnosis model to achieve the preferred fault diagnosis performance in the actual industrial scenario.In view of the above problems,this paper carries out the following research work:(1)The bearing vibration signal processing method and original feature extraction based on DTCWPT(Dual-tree Complex Wavelet Packet Transform)are studied.The original rolling bearing vibration signal is decomposed by DTCWPT,different terminal nodes are obtained and reconstructed,and the statistical parameters of reconstructed signal and the corresponding Hilbert envelope spectrum are calculated to obtain the original statistical feature set.(2)The statistical features evaluation method is studied.After the process of time-frequency analysis method,in view of the problems that the high dimension of the original feature set in which has interference and redundancy features,the PSFSC(priority selection method of features based on fault sensitivity and correlation between features)is proposed.PSFSC can evaluate two aspects of statistical features.One is the fault status sensitivity of statistical features,the releif F algorithm is used to evaluate the between-class differentiation degree of statistical features,and then the standard deviation of feature samples is calculated to represent the intra-class aggregation degree of statistical features.The ratio of the index that represents the between-class differentiation degree and the index that represents the intra-class aggregation degree is used as the index that represents the fault status sensitivity of statistical features.The other is the correlation degree between statistical features and other features in the feature set.The PCC(Person Correlation Coefficient)between statistical features and other statistical features in the feature set is calculated,and then the sum of the PCC can be obtained to represent the correlation between statistical features and other features in the whole feature set.Finally,a new feature evaluation index,FPSD(Feature Priority Selection Degree),is proposed to evaluate the statistical features.The statistical features with high FPSD are selected as the preferred features,which are used to construct feature subsets.(3)The rolling bearing fault diagnosis based on preferred features and dimensionality reduction is studied.According to the process of data-driven fault diagnosis,a rolling bearing fault diagnosis framework based on preferred features and dimensionality reduction is constructed on the basis of bearing vibration signal process based on DTCWPT,priority selection method of features PSFSC,dimensionality reduction methods(PCA,LDA,LFDA and NPE)and support vector machine classifier.Based on two kinds of rolling bearing fault data(which are from case western reserve university test bed and SQI-MFS test bed respectively),the fault diagnosis experiments under the same working condition and different working conditions are carried out.The experimental results show that the proposed PSFSC method can significantly improve the fault diagnosis accuracy under the appropriate preferred feature number,and the fault diagnosis model using PSFSC can attain the ideal fault diagnosis performance under different working conditions,which indicates that the proposed method it has the potential to be applied in practical industrial scenarios.(4)The rolling bearing fault diagnosis based on preferred features and transfer learning is studied.In view of the lack of sufficient labeled training fault data in the fault diagnosis field,as well as the distribution difference between the training samples and test samples,transfer learning method,as a new research idea in the fault diagnosis field,has attracted more attention.Based on the research of joint distribution adaption(JDA),a modified joint distribution adaption(MJDA)feature transfer learning method is proposed in this paper.Based on the bearing vibration signal processing method based on DTCWPT,priority selection method of features PSFSC,MJDA,and SVM classifier,a fault diagnosis framework of rolling bearing based on preferred features and MJDA is constructed.In order to verify the effectiveness and adaptability of the proposed methods,two kinds of rolling bearing fault data(which are from case western reserve university test bed and SQI-MFS test bed respectively)are used for experimental analysis.In the experimental case,one kind of fault data with labels is used to train the fault diagnosis model,and the other kind of fault data without labels is used as the test set.The experimental results show that the MJDA method can significantly improve the fault diagnosis accuracy,and the fault diagnosis model combined with PSFSC method can achieve ideal fault diagnosis performance when selecting the appropriate number of preferred features,which further shows that the proposed method and fault diagnosis framework have the potential to be applied in the actual industrial scenarios.According to the experimental results of the proposed two fault diagnosis frameworks,it can be seen that the priority selection method of features proposed in this paper can effectively select the features which are more conducive to fault pattern recognition from the original feature set.When the appropriate number of preferred features is selected,the fault diagnosis model can achieve the ideal fault diagnosis performance.In the process of fault diagnosis under different working conditions,the propoesd rolling bearing fault diagnosis framework based on the preferred features and MJDA has better diagnosis performance than that of the fault diagnosis framework based on preferred features and dimensionality reduction,which can further verify the superiority of MJDA trasfer learning method on improving the fault diagnosis performance under different working conditions.The paper has 50 drawings,31 tables and 184 references.