Research On Rolling Element Bearing Fault Diagnosis Methods Based On Manifold Learning

Rolling element bearing is the most commonly used key component of machinery. It is an important part of the equipment maintenance work to ensure the safe and stable operation of rolling element bearing. But the working condition of the rolling element bearing is very complex, and its speed varies greatly, the bearing way is sundry, the movement is uncertain. These factors would adversely affect rolling bearing fault diagnosis, thereby reducing the performance of traditional diagnosis methods. In this paper, the rolling element bearing fault signal is regarded as research object. The manifold learning method is combined with other modern signal processing theory to research a series of questions such as noise reduction, feature extraction, fault source separation, performance monitoring in the process of rolling element bearing fault diagnosis. The main research contents and conclusions are as follows:1. The significance of rolling bearing fault diagnosis is discussed. The features of rolling element bearing fault diagnosis technology in different development phases are analysed. Problems such as noise reduction, feature extraction, fault source separation, performance monitoring are encountered in the rolling element bearing fault diagnosis; and the research status of each issue is reviewed. The nonlinear manifold learning theory and its application in fault diagnosis is introduced on the basis of summing up the fault diagnosis technology trends of rolling element bearing.2. For rolling element bearing fault signal denoising, the dual tree complex wavelet (DTCWT) denoising method based on maximum variance unfolding (MVU) is proposed. The signal subspace of DTCWT detail signal space is extracted by MVU. The noise reduction is achieved by removing the noise subspace. The defects of conventional wavelet transform such as translational sensitivity and non-perfect reconstruction can be overcomed by the translation invariance and perfect reconstruction of DTCWT. The nonlinear structure of the high-dimensional data space can be effectively extracted by MVU manifold lgorithm; the lack of linear structure is overcomed. The both advantages are combined by DTCWT_MVU denoising method. Not only a higher signal-to-noise ratio can be obtained, but also the nonlinear impact of bearing fault signal can be obtained through DTCWT_MVU denosing method. The denoised waveform distortion is reduced. The effectiveness of this method is vertified by the emulation signal and the engineering signal.3. For extracting the fault features of rolling element bearings, the time-frequency fault feature extraction method based on tensor manifold learning is proposed. Based on the time-frequency features of HHT, the signals nonlinear tensor manifold time-frequency features are extracted by tensor manifold learning method. The time-frequency feature parameters are defined. The bearing fault samples are accurately classified by combing the tensor manifold time-frequency feature parameters with the probabilistic neural network. The intrinsic nonlinear features of high-dimensional time-frequency features can be effectively extracted by tensor manifold learning method. Compared with HHT time frequency feature parameters, the tensor manifold time-frequency feature parameters can be used to reduce the redundancy of the feature information, to more effectively classify the different types of fault samples, to reduce the iterations of neural network, to improve the fault classification accuracy.4. To separate the rolling element bearing fault source, the fault source blind separation method of rolling element bearing based on manifold learning is proposed. The multi-channel test signals are structured by EMD decomposition. The number of test signal source is estimated. The optimal test signal selection criteria are established. The best test signal is chosen by comprehensivly utilizing the kurtosis, sparsity, mutual information standard. The fault source is effectively separatated by extracting the KPCA manifold composition as the input to ICA algorithm. The optimal test signal selection in underdetermined blind separation process is solved by this method. The separation ability of ICA is enhanced by using manifold learning, so that the fault source with obvious impact feature can be separated from the single-channel signal with weak fault information.5. To deal with the performance degradation monitoring of rolling element bearing, the performance monitoring method based on manifold learning and fuzzy clustering is proposed. Sensitive band of the monitoring signal is determined by wavelet packet decomposition. On this basis, the low dimensional manifold features of signal are extracted as data samples of fuzzy clustering. Sample membership can be used as performance indicator to monitor the law of bearing performance degradation. Compared with the monitoring methods based on single feature and linear multi-features, the four stages of rolling element bearing performance degradation in the whole life cycle can be effectively reflected by this method. The uniform performance degradation law of rolling element bearing parts can be reflected. The early bearing fault can be predicted in advance.6. Bearing fault analysis and diagnosis system based on manifold learning is developed through LabVIEW and MATLAB mixed programming. The system development environment and structure scheme of hardware and software is introduced. The basic functions of the system are demonstrated. The effectiveness of the system is verified by examples.