Study On Cyclostationarity And Blind Source Separation-Based Rolling Element Bearing Fault Feature Extraction

With the rapid development of technology and industry, mechanical equipment hasbecome more and more huge, complex, high-speed, effective, and heavy-load while they mustface more and more harsh running conditions. Once they fail unexpectedly, the unexpectedfailure will increase maintenance cost, reduce production efficiency, and sometimes causesignificant economic losses, or even catastrophic accidents.Mechanical systems usually have complex structure and lots of parts. In conditionmonitoring of mechanical equipment, we usually use a multi-measurement point, multi-sensoracquisition mode. Each measurement is often a mixture of vibration signals of variouscomponents. And the transmission path of the component to the sensor is uncertain. If eachsource signal can be effectively separated, it will be helpful for exact feature extraction andstate identification. Blind separation algorithm combined with cyclostationarity can be usedfor feature extraction of rotating machinery vibration signals under strong background noise.The cyclostationary model of rolling element bearing vibration signal is taken as the basis ofresearch and development of new feature extraction methods. The contents are as follows:(1) From the viewpoint of theoretical analysis and engineering application, thebackground and significance of the selected topic are discussed. The development ofequipment condition monitoring and fault diagnosis, cyclostationary mechanism for rotatingmachinery, fault diagnosis methods based on cyclostationarity, and blind signal processing.The specific research points are decided, and the research contents of this paper areintroduced.(2) The cyclostationary theory and second-order cyclic statistics is introduced. Themathematical model of pitting faulty bearing signal is built. Based on the definition ofsecond-order cyclic statistics, researched the second-order cyclostationarity of bearing signal swith various types of fault.(3) Considering the reference signal is difficult to be obtained, taking advantage of thecyclostationarity of rolling bearings, a cyclic Wiener filter using the observed signal asreference signal is proposed. The frequency shifted version is used as input of the filter. Thecyclostationary signal is correlative with its frequency shifted version, and the stationarywhite is irrelative with its frequency shifted version. After a filter-bank, the noise will bereduced. The analysis of rolling bearing accelerated life test data shows the cyclic Wienerfilter can be used to extract weak fault feature.(4) After introducing the mathematical description of blind source separation, thepractical problems of blind separation in engineering applications are researched, includingmethods of source number estimation, noise elimination and the uncertainty of blindseparation. (5) The sources are supposed as multiple convolved faulty bearing signals. A methodbased on second-order cyclic statistics is proposed. STFT is used to transform the convolvedmixture in time domain into instantaneous mixture in frequency domain. The jointapproximate diagonalization to spectral correlation density matrix is used to obtain theseparation matrix. The observed signal is filtered by the separation filter. The separated signalis a reliable estimation of source signals. Under the assumption of cyclostationary sources, themethod exploiting the spectral coherence function and cyclic spectral density is proposed tosolve the problem of uncertainty of blind separation in frequency domain.(6) For a collected gearbox signal, the fault feature of bearings may be affected by thegearmesh frequency. A synthesized fault diagnosis method is introduced. The analysis of datacollected from the typical gearbox fault simulation test bench verified the effectiveness of theproposed separation method.