Research On Early Fault Prediction Method Of Rolling Bearing Based On Multi-Manifold Learning

As the supporting parts and transmission parts of rotating machinery,rolling bearings will affect the whole transmission system once there is a problem,and then lead to the normal operation of the equipment.Therefore,this thesis carried out the research on the early fault prediction method of rolling bearings,that is,accurate identification of faults at the germination stage,early detection of problems,prevention.Taking rolling bearings of rotating machinery as the research object,based on the idea of high dimensional manifold space,the feature extraction of early fault of rolling bearings was studied,and combined with width learning and gated cycle unit,the early fault diagnosis and prediction of rolling bearings were realized.The main research work is as follows:(1)Research on early fault feature extraction method of rolling bearing based on multi-manifold learning: In view of the fact that most of the existing feature extraction methods analyze the statistical features of samples and lack attention to the local geometric structure information,this thesis introduces the density scaling factor neighborhood adjustment algorithm on the basis of the traditional local linear embedding algorithm,and combines the sample label adaptively to construct the inner and intermanifold graphs.Then,considering that the minimum linear representation error of the local linear embedding algorithm can reflect the degree of sample aggregation to a certain extent,an adaptive local linear embedding algorithm was proposed by constructing the divergence matrix of the manifold inner graph and the manifold edge distance,which realized the early fault sensitive feature extraction of rolling bearings.The experimental comparison and analysis were carried out,and the effectiveness of the proposed method was verified.(2)The traditional deep neural network has complicated parameters,long training time and wasted computing resources,while the traditional width learning parameters need to be set manually.The particle swarm optimization algorithm was introduced to iteratively optimize multiple hyperparameters of width learning,and the optimal parameter set was used to train the early fault diagnosis model of width learning to realize the early fault diagnosis of rolling bearings.The particle swarm optimization support vector machine was selected for comparison experiment.Compared with the support vector machine,the average early fault recognition rate was 94.26% and the training time was 45.96 seconds,the average early fault recognition rate of the proposed method was97.14% and the training time was 0.004751 seconds,greatly reducing the training time.Moreover,the early fault recognition rate has been improved to some extent,which verifies the superiority of the proposed method in training time and fault recognition rate,and verifies the effectiveness of the proposed method.(3)Most of the existing fault prediction algorithms are based on typical faults or residual life prediction,and there is a lack of early fault prediction methods.Based on the traditional early fault prediction method of gated cycle unit,this paper introduces an optimization algorithm to solve the problem of immobilized and empirical gated cycle unit parameter values.Then,aiming at the traditional rolling prediction mode of gated cycle unit is susceptible to accidental factors,the prediction mode is optimized,and an improved early fault prediction method of gated cycle unit is proposed.The multi-step average method is used to increase the utilization rate of network model to historical data and reduce the influence of accidental factors,so as to improve the prediction accuracy of network model.Finally,combined with the sensitive characteristics of multi-manifold degradation of rolling bearings,comparative experimental analysis was carried out.Compared with the mean absolute error of 0.0192 and mean square error of 7.5897e-04 in early fault prediction of gated circulation unit neural network,the mean absolute error of early fault prediction of the proposed method was 0.0071 and mean square error of1.1059e-04.The error is greatly reduced.The superiority of the proposed method is verified.