Research On Rolling Bearing Fault Diagnosis Method Based On Deep Ensemble Learning

Effective fault diagnosis of rolling bearings is the vital to ensuring the safe and reliable operation of rolling bearings in industrial machinery equipment.In recent years,enabled by Machine Learning(ML)algorithms,data-driven rolling bearing fault diagnostics approaches have been steadily developed and become one of the most promising solutions in the industry.However,each ML algorithm exhibits some shortcomings limiting its applicability in practice.In this thesis,the main work is as follows:(1)Aiming at the problem that the dataset of the Rolling bearing experimental platform of Case Western Reserve University(CWRU)cannot reflect the real condition in the actual operation of rolling bearings,the experimental platform of rolling Bearing and Gear Monitor(WHUT-BGM)is designed and built.The experimental platform can simulate the working conditions of rolling bearings under various experimental conditions and collect vibration acceleration signals by replacing parts.(2)Aiming at the problem that traditional time-domain and frequent-domain features cannot accurately reflect the actual conditions of rolling bearings under complex working conditions in the field of rolling bearing fault diagnosis,a "deep features" extraction method based on Convolutional Neural Network(CNN)and Representation Learning is proposed.In order to reduce the difficulty of CNN parameter tuning,this method combines the advantages of signal processing methods to construct multi-domain CNNs to extract " deep features".At the same time,based on traditional time-domain,frequency-domain features and " deep features",rolling bearing fault diagnosis experiments are designed on multiple datasets,which verify the effectiveness of the proposed multi-domain " deep features" extraction method of rolling bearings based on CNN and representation learning.(3)Aiming at the problem of low computational efficiency and serious interference of useless features caused by excessive number of features in " deep features" dataset,a feature optimization method based on simplified Out of Bag(OOB)dataset of Random Forest(RF)in Ensemble Learning(EL)is proposed.The RF model can be used to calculate the Feature Variable Importance Measure(FVIM)index of each feature.Experimental results show that this method can effectively select the most valuable " deep features" from dataset to improve computing efficiency.(4)Aiming at the problem that assigning “equivalent cost” to different rolling bearing fault types will affect the identification of severe faults with greater harm,a fault diagnosis method based on Gradient Boosting Decision Tree(GBDT)in Ensemble Learning and an integrated strategy of Ensemble Leaning named“Non-equivalent Cost” Logistic Regression(NCLR)is proposed.Experimental results show that this method can identify the dangerous faults more efficiently and ensure the safety of equipment operation.