Research On Fault Diagnosis Of Rolling Element Bearings Based On Spectral Kurtosis And Atomic Decomposition

Rolling element bearing (REB) is one of the most widely used standardcomponents in rotating machinery. Its working state is directly related to theproduction quality and the safety of the entire unit and even the entire productionline. Therefore, it is very significant to research on techniques of bearing faultdiagnosis.Feature extraction is the critical technique for fault diagnosis. In this paper,according to the shortcomings of the exist methods for the feature extraction of REBfault, two feature extraction methods of the spectral kurtosis and atomicdecomposition are researched deeply.(1) The spectral kurtosis provides an effective tool to choice the bestnarrow-band adaptively for the classical demodulated resonance technique. However,the practicability of the method of Short Time Fourier Transform spectral kurtosis islimited due to the extensive calculation, in which two parameters are determinedsynchronously. Whilst the Protrugram technique breaks the adaptation of bandwidthselection since the bandwidth is preset before the center frequency is determined bysearching through the frequency spectrum. Therefore, based on the amplitude-modulated characteristic of the REB fault vibration signal, the technique of theenvelope positioning FFT spectral kurtosis is proposed. The contradiction betweenthe calculation and the adaption is solved by step-by-step determining the bandwidthand the center frequency. The effectiveness, adaptation and practicability of thenovel technique are verified through a comparative study of three methods of rollingbearing fault diagnosis.(2) Finite impulse response (FIR) filter fast kurtgram is able to determine thebandwidth and the center frequency fast and approximatively. However, the methodof FIR Kurtgram has certain limitations to the feature extraction of nonstationarysignals since the filter is based on Fourier transform. Therefore, in this paper, the wavelet packet (WP) filter is proposed to take the place of FIR filter in fast kurtgramtechnique. According to the broad band characteristic of fault REB information infrequency spectrum, accumulated envelope spectrum method is proposed based onthe synchronous average denoise principle. The envelope spectrums of all or part ofsub-band signals in a given level of WP are accumulated instead of selecting onenarrow-band signal, which enhances the useful information of fault REB effectivelyand improves the ability of identification to the REB fault features.(3) The dictionary constituted with the same feature type atoms cannot adapt tothe actual signals which are mixed together by a variety of complex physicalphenomena, such that the sparsity of the signal decomposition is insufficient and thephysical interpretation is difficult to be obtained. Based on the REB vibration signalcharacteristics, the mixed dictionary is constructed from the cosine packet (CP)atoms and the wavelet packet (WP) atoms. The fast CPWP hybrid atomic matchingpursuit decomposition algorithm is proposed, which improves the sparsity ofdecomposition and enhances the physical interpretability. The results of the REBfault diagnosis using the method of CPWP atomic decomposition indicates that thematching pursuit for the fault REB signal with the CPWP mixed dictionary caneffectively extract the impact components and the carrier components, whichperfectly reflects the fault features.(4) It is greatly difficult to construct the parametric atoms consistent with theactual signals which are time-varying and complex. Therefore, in this paper, basedon the dual nature of cyclostationarity and stochasticity of the REB fault signal, thenon-parametric atomic decomposition technique is proposed due to the relaxedconditions to the atomic construction and the approximate expression to the signal inthe matching pursuit method. The non-parametric atomic decomposition methodextracts the feature waveforms from the fault signal, which are used to construct thenon-parametric dictionary. The superiority of the new approach is studied fromseveral aspects of matching, sparsity and frequency resolution. The measured signal of the fault REB is analyzed by three methods. Compared with the cyclic spectrumdensity methods and envelope demodulation method, the validity of the method ofnonparametric atomic decomposition is verified.