Research On The Fault Diagnosis Method Of Rolling Element Bearing Based On Cyclostationary Signal Processing

Rolling element bearing is one of the most widely used components in almost all kinds of machinery. Its working condition has influence on product quality and working safety in industry directly. So it is very significant to research on bearing condition monitoring and fault diagnosis techniques, especially for bearing early fault.Some techniques of condition monitoring for rolling element bearing have been analyzed in the thesis. Since symmetry or approximately symmetry of working parts of a bearing and its uniquely periodical rotating working mode, vibration of a destroyed bearing generally exhibits strong periodicity. These periodical pulses bring amplitude modulation (AM) characteristic. Considering the above reasons, a new signal processing method for bearing condition monitoring is studied in this thesis.Firstly, first-order cyclic statistics, named Cyclic Mean (CM) value, is studied. An example result has been obtained by cyclic mean for the amplitude modulation signal, which proved the signal exhibit cyclostationarity. The results of CM analysis and Synchronous Average (SA) analysis are compared. It is pointed out that the two methods possess similar performance for identifying the first-order periodical feature of the signal. However, their basis theories are different, the CM is based on sine-wave generating algorithm, but the SA is just based on the average estimate. A modified Cyclic Temporary Synchronous Average (CTSA) method is put forward. The CTSA analysis method can exhibit the first-order multi-periodical characteristics of the signal much more clearly. Our research also indicates that cyclic mean does not possess demodulation function.A further conclusion is introduced that Cyclic Autocorrelation Function (CAF) possesses demodulation capability. Comparing CAF with envelope demodulation and square demodulation, it is proved that the three methods have a similar performance on extracting feature frequencies of an AM signal. However, the three methods also have essential difference. CAF can obtain the periodical information for a cyclostationary signal by nonlinear transform and cyclic factor. It demodulates the AM components to cyclic frequency axis directly. Envelop demodulation technique only extracts AM frequencies from the original signal, and does not obtain any information about carrier wave. Square demodulation technique is just a special part of the CAF when time lag is equal to zero.Further study focuses on the slices of amplitude of CAF at different cyclic frequencies. The results show that the slices of the module of CAF can extract different information corresponding to different cyclic frequencies. At some certain cyclic frequencies, the module of CAF even can extract pure feature frequency. Several examples have been analyzed and common rules are summarized.Spectrum Correlation Density Function (SCDF) is studied in this thesis. The time smoothed and frequency smoothed cyclic periodogram are imported to obtain SCDF. Slices of SCDF for some cyclic frequencies are analyzed, and the similar rules are obtained as CAF. There are many kinds of transform modes of SCDF. The symmetrical and nonsymmetrical transformations are commonly used. The result of SCDF via a usual transformation is deducted. The result shows that different transform will affect information expression for SCDF. However, the essences which the SCDF provided are the same.Degree of Cyclostationary (DCS) is discussed, and a special DCS definition for engineering application is provided. It is a scale for the intensity of periodical information for non-stationary signal. A viewpoint is referred that corresponding signal processing technique should be adopted to analyze signal with different DCS.Higher-order cyclic statistics (HOCS) theory is also discussed. Some performances have been compared with second-order statistics, higher-order statistics and higher-order cyclic statistics. The conclusion is that HOCS has better anti-interference performance. The slice analysis method of SOCS is extended to HOCS domain, which can express the results of HOCS more clearly. However, HOCS has some limitations, such as the implement method is complexly and the calculation data is too long.A new cyclostationary characteristic strengthening method of weak cyclostationary signal based on Morlet Continuous Wavelet Transform (CWT) is advanced. This method uses the characteristics of Morlet Wavelet and the advantage of CWT, according to the minimum of entropy principle and the threshold maximum likelihood estimate method, and at last can strengthen the cyclostationary characteristics of the weak cyclostationary signal.The traditional vibration model of fault rolling element bearings is modified and gets a new cyclic model. The new model uses three random time-varying parameters to simulate the time-varying of the amplitude, the period of the impulse and the response of the system to this impulse. These improvements make the new model approach to the real work condition of rolling element bearings much more closely.Experimental research is well done at last. Several different kinds of rolling element bearing local faults are investigated. The vibration data are acquired and analyzed, the results verifies the accuracy of corresponding theoretical research in the present study and demonstrates its availability and feasibility in engineering practices. An intelligence diagnosis system of rolling element bearings based on cyclic statistics and BP Neural Networks is established, which provides a new kind of method to solve the intelligence diagnosis problem of rolling element bearings.