Research On Health Status Identification Of Rolling Bearing Under Data Class Equilibrium Problem

Rolling bearings,as key components in rotating machinery,are widely used in wind turbines,CNC machine tools,industrial robots and other equipment.Because rolling bearings are in high temperature,variable speed and variable load for a long time,which leads to rolling bearings easily damaged,and then cause mechanical equipment failure,which may lead to production stoppage in serious cases.Therefore,in order to improve the reliability of rolling bearings,it is necessary to monitor and identify their health status in a timely manner.In recent years,rolling bearing intelligent health status identification method has become an important means to ensure the safe work of machinery and equipment under big data by automatically extracting the deep features in the signal to diagnose the health status of the machine.A deep learning-based classbalanced rolling bearing health status identification method is proposed for the situation where typical fault information is abundant and the categories are balanced between healthy and faulty samples.At the same time,because in the actual working condition,the equipment operates normally for a long time,it is easy to cause the samples obtained in the normal state are abundant while the samples obtained in the fault state are sparse,which results in the category imbalance between the healthy samples and the fault samples and causes the model recognition accuracy to decrease.A deep learning-based method for identifying the health status of class unbalanced rolling bearings is proposed for this problem.In this paper,we investigate the class balance and class imbalance situations existing between faulty and healthy samples in rolling bearing health status identification,and give the corresponding bearing health status identification method based on deep learning,as follows.(1)When rolling bearing failure data and health data categories are balanced.The existing deep learning-based bearing health status identification model has the following problems.The first one is that the learning rate of the optimizer often adopts default values or manual adjustment,which easily leads to low adaptive performance of the diagnostic model and cannot adapt to the variable operating conditions of bearing health status identification.The second,Convolutional neural network reduces the dimensiona-lity of the input image by max pooling or mean pooling,the pooling operation will lead to the loss of location information between the data and the merging operation may change the local trend of the signal,which leads to the signal being misclassified.This problem can be well solved by expanding the perceptual field by making the convolutional kernel size larger.However,enlarging the field of perception by this method will lead to poor performance of the network model and an increase in the overall number of parameters and computational cost.The third,Adam is widely used as a representative algorithm of gradient descent method,but due to the setup of Adam,some small batches provide large gradients,and although these large gradients are rich in information,their influence will disappear quickly due to exponential averaging,which leads to poor convergence of the diagnostic model.An adaptive resource allocation deep neural network(PDC-LR-HCNN)consisting of an improved extended convolution,lookhead-radam(LR)optimizer,and superband regulator is proposed for the above-mentioned problems of deep learning models when the rolling bearing fault data and health data categories are balanced.The proposed PDC-LR-HCNN neural network model is validated with the experimental dataset of CWRU and XJTU-SY bearings,and its accuracy can reach more than 90%,and the classification accuracy can be converged within 10 epochs.(2)When rolling bearing failure data and health data categories are imbalanced.In this case,the health status identification model used for the class balance case is not applicable.The supervised transfer learning-based Conv Next rolling bearing health status identification method TConv Ne Xt is proposed,which reconstructs the rolling bearing dataset into a balanced dataset by the Synthetic Minority Oversampling Technique(SMOTE)method.Afterwards,transfer learning is used to enable the TConnectv Ne Xt network model to grasp some of the weights needed to discriminate the composite fault information of rolling bearings,followed by the conversion of the one-dimensional signals into RGB images using Hilbert transform to input the model and train the remaining weights of the model.Afterwards,transfer learning is used to enable the TConv Ne Xt network model to master some of the weights needed to discriminate the composite fault information of rolling bearings.Next,the model’s remaining weights are trained by converting the one-dimensional signal into a 2D image input model using Hilbert transform.The diagnostic model does not rely entirely on the data monitored at the fault site for training,which greatly reduces the reliance of the health status identification model on the on-site data.Finally,the trained TConv Ne Xt network model is used for health status identification and visualized and analyzed using the Grad-CAM method.The experimental results show that the TConv Ne Xt network model has a high diagnostic accuracy and can better meet the requirements of rolling health status identification in the case of class imbalance.(3)The proposed health state identification method in the case of class imbalance is improved for the problem of data overlap in the process of data resampling and the problem that the bearing health state learned in advance by supervised transfer learning is relatively single.The Edited Nearest Neighbors(ENN)undersampling method is used to eliminate the problem of data overlap caused by using the SMOTE.Using secondary transfer learning caused the model to perform the first transfer learning with the full life cycle bearing dataset as the source domain and the Case Western Reserve University bearing dataset as the target domain,followed by the second transfer learning with the laboratoryderived dataset to train the model.Compared with supervised transfer learning,secondary transfer learning can acquire more bearing health state features in advance,which in turn can further improve the accuracy of model health state recognition in the class imbalance case.