Research On Noise Reduction And Fault Feature Reduction Of Bearing Vibration Signal

With the continuous development and progress of science and technology and manufacturing engineering,rotating machinery equipment is facing the trend of highspeed,precision,complexity and intelligent development,and the requirements for stable and reliable operation of equipment are also constantly improving.Rolling bearing is an important part of rotating equipment.The smooth operation of industrial equipment depends on the normal operation of bearing.Its slight defects may lead to the failure of the whole mechanical system.Therefore,the detection and subsequent diagnosis of bearing state is an important subject.In a word,this paper takes rolling bearings as an example to study three key issues: first,noise reduction of rolling bearing vibration signals;Second,the reduction of high-dimensional features of rolling bearings;Third,rolling bearing fault type classification.The principal contents of this paper including:(1)About noise disturbance of vibration signal of rolling bearing,variable mode decomposition(VMD)combined with wavelet semi soft threshold is selected to denoise the rolling bearing vibration signal,parameter selection problem in VMD decomposition,optimization using the algorithm of whales.Aiming at the problem of high frequency noise,a VMD denoise method based on mutual information is proposed,The high frequency parts is selected according to the mutual information criterion from intrinsic mode function(IMF)after VMD decomposition,And this part is denoised by wavelet semi-soft threshold,Finally,the low frequency and high frequency after noise reduction are combined to eliminate the interference of noise on vibration signal analysis.Through the simulation analysis of bearing inner ring and outer ring faults,the effectiveness of this method is verified.(2)Sixteen time domain features,four frequency domain features and three entropy features were extracted from the denoised signal to form the highdimensionality collection in the mixed domain.In view of the information redundancy of high-dimensionality solicitation,easy to cause "over-fitting" and low computational efficiency,a manifold learning algorithm is proposed to transform the high-dimensional features into low-dimensional space.The local linear bedding(LLE)algorithm is mainly improved.To solve the problem that euclide distance cannot measure the relative distance of two points on the streamline and sample points need to be evenly distributed,a uniform Geodesic rank-order algorithm is proposed.It improves the original algorithm’s ability to preserve the manifold structure of sample space.The results show that the improved algorithm improves the separability of the reduced data.(3)This paper takes the standard data set of bearing fault diagnosis recognized in the industry as the basis: Case Western Reserve University data set as an example to test the feasibility and effectiveness of VMD noise reduction based on mutual information and improved LLE algorithm for dimensionality reduction.Two groups of experiments with different fault types and fault sizes of rolling bearings were designed.Considering that the selection of parameters in the reduction process would affect the reduction results,a grid search method was proposed to select the number of adjacent points k and reduced dimension d of the improved LLE parameters.Experimental results show that this method can keep original signal information while filtering highfrequency noise.The data after the improved LLE dimension reduction is significantly improved compared with the Fisher measure before the improvement,and is more easily classified by SVM.