Research On Machine Learning Based Fault Diagnosis Method For Rotating Machinery Bearings

Rotating machinery plays an important role in industrial production and is one of the most widely used types of equipment in mechanical equipment,including generators,blowers,electric motors,steam turbines,pumps,etc.Their operational status directly affects production efficiency,safety,and cost.In the fault detection of rotating machinery,most of the faults are caused by key components.Among the many components of rotating machinery,bearings are the most frequently faulty component.Therefore,the research on bearing fault diagnosis technology has important theoretical value and practical significance.This paper takes bearings as the research object,and focuses on typical problems such as Fault diagnosis of bearings under single operating conditions,fault diagnosis under multiple working conditions and fault diagnosis under variable working conditions.The following work is carried out:(1)Two methods have been proposed for diagnosing bearing faults under single operating conditions,one for diagnosing the fault location of bearings and the other for diagnosing bearing faults with different degrees of damage.The bearing fault diagnosis method based on correlation analysis and grid search algorithm optimization using support vector machine the first to propose using SVM classification results for correlation analysis,and cleverly introduces the idea of layering in model building.The bearing fault diagnosis method based on sparse classifier uses all training samples as dictionaries to sequentially solve sparse coefficients for each test sample.Then,filter out the corresponding parts of each category in the sparse coefficient and solve for the category error.Finally,the fault classification of bearings is carried out by comparing category errors.Through the validation of the CWRU bearing dataset,the results show that both proposed methods can effectively diagnose the fault location without through fault signal denoising and feature extraction processes.Moreover,the bearing fault diagnosis method based on sparse classifiers shows good performance in fault diagnosis of different degrees of damage.(2)Two methods have been proposed for bearing fault diagnosis under multiple working conditions,which are respectively used to solve bearing fault diagnosis in the presence of significant noise in bearing fault signals and small sample situations.The bearing fault diagnosis method based on signal image mutual mapping and sparse representation innovatively converts preprocessed fault signals into grayscale images and uses dictionary learning for block denoising before restoring them to fault signals.The bearing fault diagnosis method based on correlation coefficient weighted signal denoising and entropy feature fusion proposes for the first time to apply the filtered correlation coefficient values as weights to the corresponding intrinsic mode functions to enhance the fault signal.Secondly,after calculating and fusing the refined composite multiscale dispersion entropy(RCMDE),refined composite multiscale fluctuation dispersion entropy(RCMFDE),refined composite multivariate generalized multiscale fuzzy entropy(RCmv MFE),refined composite multivariate generalized multiscale sample entropy(RCmv MSE),and multiscale permutation entropy(MPE)of the signal,the Fisher classifier is used for fault diagnosis.Finally,the effectiveness of the proposed method is measured using the newly defined parameter bias and diagnostic accuracy.Verified using the CWRU dataset and Paderborn dataset,the results show that compared with some existing methods,the proposed two methods achieve higher diagnostic accuracy in 12 single operating conditions and over 30 multi working condition experiments.Moreover,compared to artificial damage experiments,the proposed method performs better in real damage experiments and has better practical applicability.(3)A bearing fault diagnosis method based on multi feature fusion and improved weighted balanced distribution adaptation(IWBDA)is proposed for bearing fault diagnosis under variable operating conditions.The method first selects and fuses 13 frequency features and 4 entropy features with complementary information to characterize the source domain fault signal and the target domain fault signal.The entropy features include multiscale distribution entropy(MDE),RCMDE,RCMFDE,and RCmv MFE.Afterwards,a new progressive parameter selection method is proposed based on the characteristics of weighted balanced distribution adaptation(WBDA)to solve the problem of unstable performance of WBDA caused by parameter changes.Finally,the source domain fault features and target domain fault features fused by the proposed IWBDA transfer are used,and KNN is used as a classifier to determine the fault category.The CWRU dataset is used for validation,and the experimental results shows that the proposed method achieves good fault diagnosis performance in 12 different working conditions.In addition,comparative experiments also demonstrate the effectiveness and stability of this method.In summary,this paper starts from machine learning based bearing fault diagnosis technology,with the goal of achieving better fault diagnosis results,and conducts a series of research on bearing fault diagnosis methods under different working conditions,aiming to provide new ideas and methods for the development of bearing fault diagnosis technology and provide more reliable guarantees for the stable operation of industrial production.