Research On Improved Local Linear Embedding Method For High Dimensional Feature Reduction Of Rolling Bearing Data

In recent years,to conduct intelligent fault diagnosis on rolling element bearigns,the general procedure is to extract some features from their vibration signals and then use some machine learning methods to determine their healty states.On the one hand,the condition monitoring of bearings provides a large amount of data for analyzing their internal states,which is useful for their accurate fault diagnosis.On the other hand,such huge data also increases the difficulty of bearing analysis and diagnosis.The corresponding problems include rundundant information,confilict conclusions,high computation cost,and so on.Therefore,it is ergent to study how to dig up the implicit information in these data and then obtain the key feature sets for the following diagnosis.In this thesis,the local linear embedding(LLE)method is introduced and some impoved methods are then proposed to improve the feature reduction performance of the LLE and solve its main problems.Main contents in this research are summarized as follows:(1)In order to keep a good balance between better feature reduction and higher diagnosis accuracy,an improved an improved local linear embedding method is proposed.By changing the calculation method of the distance between two sample points,the manifold structure of the sample space is maintained,and its feature reduction performance of the algorithm is also improved.The improved LLE method is applied to processing of Iris data and bearing data and also compared with three methods,including the common LLE,the Mahalanobis-distance based LLE,and the uniform-distance based LEE.Two experimental results demonstrate the effectivenss of the proposed LLE method.(2)Considering the noise sensitibity of the LLE method and the corresponding influence on its feature reduction performance,a hybrid noise reduction method is presented.In this method,the variational mode decomposition(VMD)and the genetic algorithm(GA)are combined for signal preprocessing,the former of which is used to decompose the analyzed signal and remove the noise and other irrelevant signals,and the latter of which is used to define appropriate fitness function based on the relationship between the penalty factor in VMD and the kurtosis value of the filtered signal,and then search the optimal penalty factor.The experimental results indicate that the GA-based VMD noise reduction method can remove most of noise in the analyzed signal and avoid its unexpected influence on the feature reduction.(3)To assess the feature reduction performance of the improved LLE method,a multi-index assessment method is proposed.In this method,three measures,i.e.the classification accuracy,the similarity and the autocorrelation between the original data and the data representation after the feature reduction,are used to comprehensively evaluate the results of the improved LLE method.After applying to the bearing data,the selection of the reduction rate can be determined and then obtain statisfied results for feature reduction,which also keeps a good balance between less but better features and higher classification accuracy.The improved methods mentioned above are demonstrated to be effective for intelligent analysis of large-scale bearing data,so that they can be applied to the health monitoring and fault diagnosis of bearings.Meanwhile,the proposed results and corresponding results are also a good reference for intelligent diagnosis of bearings and other rotating machine components in other large equipment.