Fault Prediction Of Rolling Bearings Based On Partial Least Squares For Extreme Learning Machines

Rolling bearings are one of the important parts of rotating machinery,and the stability of their running state is directly related to the working performance of the overall equipment.Effectively predicting the fault location of the rolling bearing and correctly identifying the degradation stage of the rolling bearing performance are of great significance for reducing the incidence of dangerous accidents and the health management of the equipment.Aiming at the problem of rolling bearing fault prediction,the research on signal noise reduction,feature dimension reduction and fault prediction is carried out step by step.The main research contents are as follows:Firstly,an Empirical wavelet transform signal noise reduction method based on Normal distribution is proposed.In view of the traditional noise reduction methods,such as window function fixed and pattern aliasing,the Empirical wavelet is used for signal noise reduction,and the Normal distribution interval estimation method is used to overcome the problem that the original boundary calculation is complex,slow and affected by manual decision,realize adaptive division of signal spectrum;the difference between the dimension and the dimensionless in the vibration process is analyzed,and the noise reduction effect of the Normal distribution-Experience wavelet transform is verified by the simulation signal and the open source data.Secondly,a dimensionality reduction method combining Isometric Feature Mapping with Fuzzy C-means is proposed.The topology stability of the traditional Isometric Feature Mapping algorithm is susceptible to the size of the neighborhood,and the residuals that can evaluate the retention effect of the structure information are optimized.When the characteristics of the target data are generally described by fewer variables,the Fuzzy C-means is used for clustering to maintain the similarity of the manifold data in the high-dimensional and low-dimensional spaces.Through the comparison of simulation experiments,the dimensionality reduction effect of the equidistant mapping combined with the Fuzzy C-means is verified.Thirdly,the Extreme Learning Machine fault prediction model based on Partial Least Squares is established.Aiming at the problem that the network weight and the number of hidden layer nodes of the traditional Extreme Learning Machine are adjusted by experiments and are subject to human factors,the principal component number and the loading weight of the Partial Least Squares method are improved.At the same time,to better fit the data,select the activation function Softmax instead of Sigmoid.By comparing different prediction methods,the method is verified to have higher prediction accuracy.Finally,set up the test bench,collect data and test verification.Set up a rolling bearing vibration test bench and collect the required test data.The single-row angular contact ball bearings with different cracks in the inner and outer rings are taken as test objects,and the method is used to predict the fault according to the specific conditions of the signal.The comparison of test results shows that the method can accurately predict the fault location and different degrees of faults.