Research Of Rolling Bearing Fault Diagnosis Based On Coaxial Vibration Feature Fusion

Rolling bearing is the core component of rotating machinery.Once it fails,the equipment will be paralyzed,resulting in economic losses.Therefore,the research of fault diagnosis for rolling bearings is a significant subject.Most of the fault diagnosis researches focus on vibration signal analysis for the reason that the acquisition of vibration signals is easier,and the vibration signal contains rich information that can characterize the state of rolling bearings.However,the vibration signal of rolling bearing is interfered by complex noise.The noise reduction and feature extraction of vibration signal is an important part of fault diagnosis.Meanwhile,with the increase of the number of sensors in mechanical system,the vibration signal sources of rolling bearing become more and more abundant.The fusion analysis and intelligent diagnosis of multi-source data become the key research direction of rolling bearing fault diagnosis.In addition,although the amount of rolling bearing vibration data is large,there are few bearing fault samples in the industrial field,and the bearing vibration data generally have the problem of class imbalance.The research on the diagnosis model under the small sample of rolling bearing fault is the essential solution to improve the application value of the algorithm and solve practical engineering problems.Focusing on major research directions in the field of rolling bearing fault diagnosis abovementioned,a research idea for rolling bearing fault diagnosis based on the feature fusion of coaxial vibration signal is proposed in this thesis.Aiming at the problems of complex noise interference of vibration signals,multi-source information fusion and class imbalance,the following researches are carried out:(1)Since the vibration signal is disturbed by complex noise,a noise reduction algorithm for rolling bearing vibration signal based on K-adaptive variationnal mode decomposition(K-adaptive VMD)is investigated.The developed algorithm adaptively determines K by the abnormal decrease of the central frequency of intrinsic mode function(IMF),so as to obtain the number of modes as large as possible and make the frequency band of the components more concentrated.The coefficient with prior knowledge is proposed,which can effectively suppress the noise in vibration signal of rolling bearing under the interference of complex noise.(2)Based on the properties of coaxial bearing vibration mutual transmission and vibration signal coupling,a rolling bearing fault diagnosis algorithm of coaxial vibration feature fusion(CVFF)is proposed,which realizes the full use of multi-sensor monitoring data of rolling bearing.Firstly,radial basis function fuzzy entropy(RBF-Fuzzy En)vibration feature extraction method is introduced.The improved method uses RBF as fuzzy function,and develops a new method to determine the IMFs selecting index and fuzzy entropy parameters according to the distribution characteristics of the maximum absolute distance in the calculation of fuzzy entropy.Secondly,a coaxial vibration feature fusion model is proposed.Through multi-kernel learning support vector machine(MKL-SVM),the weights of kernel function for vibration feature from different sensor are adaptively obtained,and coaxial vibration feature fusion is realized by the summation of the weighted kernel function mapping.(3)A multi-scale kernel support vector machine(MSK-SVM)based on swarm intelligence optimization algorithm is proposed to solve the problems that the kernel function type of MKL-SVM is single and the kernel function parameters need to be set in advance.On the basis of common kernel functions,Morlet,Marr and DOG wavelet kernel functions are introduced.Through the improved grey wolf optimizer(IGWO),the optimization is carried out within the number of kernel functions and the scope of scale parameters.The global optimization ability of IGWO and the sparsity of multi-kernel learning can realize the adaptive selecting of kernel functions.MSK-SVM realizes the adaptive optimization of kernel function type,scale parameters,weights and other model parameters of MSK-SVM.The richness of kernel function increases the diversity of diagnostic models.(4)In view of the fact that there are many normal samples and few fault samples of bearings in the industrial field,the star generative adversarial network for sample generation(Star GAN-SG)model is studied.Combined with the CVFF algorithm,a rolling bearing fault diagnosis model under class imbalance is proposed.Star GAN-SG takes the noisy real fault samples as input.Through the adversarial training of generator and discriminator,Star GAN-SG can generate data with the same category as the input real samples,and realize multi-domain generation of multiple datasets through a single network.This model solves the problem of class imbalance of rolling bearings to a certain extent.Based on the training set composed of real and generated samples,a coaxial vibration feature fusion algorithm is constructed to realize the rolling bearing fault diagnosis under the class imbalance.