Fault Diagnosis Of Rolling Bearings Based On Improved Empirical Mode Decomposition And Multiple Feature Selection

Today,with the rapid development of science and technology and modern industry,largescale mechanical rotating equipment is widely used.On the one hand,with the increasing degree of integration,intelligence and networking,the mechanical structure is becoming more and more complex,which virtually increases the possibility of fault of mechanical equipment.On the other hand,people's requirements for the service life and quality of mechanical rotating equipment are constantly improving.Therefore,it is of great significance to predict,monitor and diagnose the equipment faults in time to ensure the reliable operation of the mechanical rotating equipment,improve the service life of the equipment and reduce the fault losses.Rolling bearing is key components of industrial machinery rotating equipment,which requires high reliability and safe operation.However,due to the heavy load work in a harsh environment for a long time,mechanical fault often occur,affecting industrial production,increasing costs and maintenance time.Therefore,the safety and reliability of rolling bearings is particularly important for industrial machinery.The timely diagnosis of rolling bearing faults is the key to ensure safety and reliability.Therefore,the fault diagnosis of rolling bearing has great research value and significance.The main work and innovative content of this paper is as follows:(1)Aiming at the problem that it is difficult to characterize and identify the signals of rotating machinery shafting systems,the vibration signal processing technology based on empirical mode decomposition(EMD)is studied,The advantages and disadvantages of the method are analyzed,and the characteristic information of shaft vibration signal is extracted by using this technology.For its inherent defect of end effect,a method based on nearest neighbor extreme value comparison is proposed to determine the extreme value distribution at the end.This method is based on the extreme point near the end point and the end point itself.It can lock the local extremum at the end point simply and quickly without complex algorithm and prediction model.It effectively improves the signal distortion caused by the end effect and the accuracy of EMD signal processing is improved.This method is applied to the feature extraction of the shaft system signal to complete the fault feature extraction of the rolling bearing.(2)Using neural network to extend data to deal with end effect has been used to some extent,but because of the large training input samples,the cost of learning time is high.On the other hand,because the predicted value exists in the input sample,the predicted value is used as the real value to predict again,which increases the error of prediction results.Therefore,BP neural network extreme value prediction method is proposed to improve this problem.The BP neural network extreme value prediction method uses the trained network model to directly predict the extreme values at the end point at one time without the need to extend the prediction of a large amount of data,which not only improves the real-time performance,but also guarantees the authenticity of the predicted data.The real vibration fault signal is used as experimental data to verify the modified method.The experimental results show that the improved method has a better effect on the end effect and ensures the authenticity of the EMD of the signal.(3)The feature extraction of time-frequency signal by EMD produces a lot of invalid features,which seriously interferes the representation accuracy of effective features on faults and affects the timeliness of fault identification.For a variety of fault categories,especially when there are many fault categories,the effectiveness of feature recognition is difficult to measure directly and accurately.In addition,after using some algorithms for feature selection,there are often redundant problems caused by the correlation between features that cannot be measured.Therefore,in order to make full use of the advantages of EMD feature extraction and eliminate the invalid redundant information,Targeting Feature Selection(TFS)model is proposed.In this method,Firstly,the complete set of fault features is obtained by analyzing the shaft system signals through empirical modal decomposition,Then,according to the idea of mutual difference,the fault sample processing space is constructed,and the mutual difference between the two fault categories is analyzed.Combined with the filter method,the features that can represent the difference are mined,and the corresponding effective feature set is obtained.These features have strong correlation with the two fault categories.Finally,by using heuristic search method and establishing heuristic search rules,we can find the feature combination which can reduce the coupling of other fault groups to the greatest extent in each effective feature set,so as to ensure the effect of multiple fault identification.Combined with the best first search strategy(BFS),the search time consumption is reduced.This method not only guarantees the representation ability of the features between any fault categories,but also the effect of each feature in the optimized feature subset on global fault identification is improved to the greatest extent.At the same time,because the feature is selected from different effective feature sets,it avoids the redundancy caused by the correlation between features,and reduces the feature dimension to the greatest extent.(4)In feature selection,the series of Relief algorithms can filter out the features that have strong correlation with the category,but they do not consider the correlation between the features.There are inevitably redundant features in the selected features,which need to further reduce the features with similar fault classification ability.Therefore,a feature selection method based on dynamic weight combination of Relief and Relief F is proposed.Firstly,the weight values of features with two fault categories are obtained by using the Relief algorithm,and the feature subsets of each two fault categories are selected according to the set threshold value,and the contribution rate of features to all two fault categories is obtained.Then,the weight values of features with various fault categories are obtained by using the Releif F algorithm,and the threshold value is set.Finally,the rule of dynamic combination of weights is established.According to the characteristics,the weights and contribution rates of all the two fault categories are combined with the weights of various fault categories to make a comprehensive judgment,and the features that meet the conditions are screened out,and the features with strong correlation between the features are removed,so as to obtain the final feature subset of multiple fault categories.This method has the advantages of Releif F series algorithm,at the same time,it adds the judgment conditions of feature selection,and makes comprehensive judgment based on the classification ability of features to all two fault categories.It not only eliminates the features based on the threshold value of Releif F feature selection,but also solves the problem that the correlation between existing features can't be measured,resulting in the redundancy of feature subsets,so it can be completely completed Feature selection task.