Research On Fault Diagnosis Method Of Rolling Bearing Based On Deep And Transfer Learning

Rolling bearing is a key component of rotating machinery equipment.In order to ensure the safe and reliable operation of rolling bearings,the real-time condition monitoring and fault diagnosis of rolling bearings must be carried out.It is difficult to construct the optimal feature set which can accurately characterize the health status of rolling bearings for traditional fault diagnosis methods,resulting in the decline of recognition accuracy of diagnostic model.In this paper,Empirical Mode Decomposition theory,Deep Learning theory and Transfer Learning method are applied to the fault diagnosis of rolling bearings and an intelligent and general fault diagnosis method is provided.This paper mainly studies the following aspects:(1)Due to the non-linear and non-stationary of rolling bearings vibration signals,the fault diagnosis of rolling bearings based on Improved EEMD and fault sensitive IMF component is studied.The principle and algorithm of EEMD and the selection method of parameters of Improved EEMD algorithm are analyzed.Combining the correlation coefficient of IMF component and its vibration signal,the kurtosis index of IMF component and the energy ratio of IMF component,the fault sensitive IMF component is automatically extracted by the comprehensive evaluation index.The fault sensitive IMF component is reconstructed,and the envelope spectrum of reconstructed signal is obtained by Hilbert transform and FFT to extract the fault feature information of rolling bearing.The effectiveness of Improved EEMD algorithm and automatic extraction algorithm of fault sensitive IMF component in rolling bearing fault diagnosis is verified by experiments.(2)Because the traditional fault intelligent diagnosis methods have the disadvantages of low diagnostic accuracy and poor robustness,the intelligent diagnosis method of rolling bearing fault based on Improved EEMD and DCNN is studied.Firstly,the Improved EEMD algorithm is used to transform adaptively the vibration signal into the IMF matrix to enhance the health status feature information of rolling bearings.Then,DCNN is used to extract adaptively discriminative fault features from IMF matrix to replace the traditionally designed feature set with poor representation performance.Finally,the fault features are identified by Softmax classifier.The diagnostic performance and anti-noise effect of the method are validated by two parts of data of artificial fault and real fault for rolling bearing.The design of the convolutional neural network architecture and key hyperparameter is analyzed in detail,and the t-SNE algorithm is further used to visualize the high-dimensional features extracted from the intermediate layer of DCNN.The mixed fault data set of rolling bearing and gear verify the universality of the intelligent diagnosis method.(3)Because the vibration signal data of rolling bearing under different working conditions has different distribution or different feature space,the recognition accuracy of the fault intelligent diagnosis model of rolling bearing is degraded.Therefore,the fault diagnosis method of rolling bearing based on DCNN and TL is studied.The DCNN is trained with sufficient and labeled samples in a working condition;the fine-tuning layer and transferring layer of DCNN are determined by experiment,then the parameters of the fine-tuning layer of DCNN are finetuned with a small number of labeled samples in another working condition to monitor fault status of rolling bearings under different working conditions.The validity of the method is verified by the fault transfer diagnosis experiment of rolling bearing under different working conditions.