Fault Diagnosis Of Rolling Bearings Based On Data-driven Model

Bearing is one of the important components of mechanical equipment in many fields.In industrial production,the real-time monitoring of bearings in operation can not only prevent accidents,but also keep the equipment stable and high-speed.There are some problems in the traditional bearing fault,such as the single way of extracting time-frequency features,which makes the signal not fully observed and the existing manifold learning dimensionality reduction algorithm sensitive to the noise mixed in the signal acquisition.In this paper,a tensor model of bearing vibration signals is established,and the key information of vibration signals in different subspaces is fused to overcome the shortcoming of a single feature extraction method of traditional fault diagnosis methods.Based on the tensor model of bearing vibration signals,this paper carried out data-driven rolling bearing fault diagnosis,improved the local linear embedding algorithm,and implemented local difference enhancement of initial features from tensor decomposition,smooth spline fitting and further reduction of initial features.The specific research contents of this paper are as follows:Aiming at the problem that a single signal processing method cannot fully observe the signal and fully extract features,tensor modeling is used to fuse different subspace information.The characteristic of bearing fault vibration signal is that its vibration frequency is mainly distributed in high frequency range.In order to reduce the interference of the low frequency part of the signal and enhance the signal of other frequencies,the following subspaces are established,including the reconstruction of wavelet space and phase space of Gauss high pass filter of bearing signal.The DWT-PSR model of bearing vibration signal constructed with them is used to improve the proportion of different frequency signals.Aiming at the problem that the high sampling rate of bearings operating at high speeds results in a large amount of redundant information in the acquired signals,which increases the difficulty of subsequent identification,this paper uses high-order singular value decomposition(HOSVD)to decompose the tensor model and get the initial features.In order to eliminate the interference of the meaningless extremum in the initial feature signal,the fitting of smoothing spline is proposed to process the feature signal,and the optimal classification point index is proposed to measure the difference between different types of information of the feature signal.Applying them to the standard bearing data set of Case Western Reserve University(CWRU),it is proved that the initial characteristic curve fitted by smoothing spline not only improves the accuracy,but also increases the local differences among different types of features.For the problem of processing a large amount of existing offline data and real-time and fast processing of new samples,considering that the initial features still have a certain scale,this paper proposes a local difference linear embedding algorithm(LDMLLE)to further reduce the dimensions of the initial features.On the basis of the optimal classification point index of the initial feature,the algorithm further obtains the local difference matrix of the initial feature and weighted sum it with the nearest neighbor matrix of the local linear embedding algorithm,so that the global distance and the local nearest neighbor can be used as the dimension reduction standard at the same time.In the LDMLLE algorithm,the generalized regression neural network model is introduced to get the LDMLLE-GRNN incremental algorithm,which can realize the online processing of bearing fault information feature extraction.Compared with other dimensionality reduction algorithms,the incremental algorithm proposed in this paper has a good real-time processing ability in the case of only a small number of prior samples,and can obtain significant low-dimensional characteristics of incremental signals.Combining the above research results with support vector machine(SVM),a complete set of bearing fault diagnosis method is given.The method is applied to the bearing fault simulation test platform in the laboratory for verification.The experimental results show that the LDMLLE algorithm has better dimension reduction effect by considering the global dimension reduction and referring to the local differences of the optimal classification points.The incremental algorithm can reduce the dimension of the new samples accurately by using a small number of prior samples,which has the advantage of real-time processing.