Methodology Of Fault Diagnosis Of Rotating Machinery Via Multiscale And Nonlinear Condition Feature Enhancement

With the research aim of precision diagnosis of rotating machinery, this paper addresses on the application studies of multiscale and nonlinear analysis for machine condition monitoring and fault diagnosis. Through analyzing the research literatures of the developing time-frequency-analysis-based multiscale analysis methods and two typical nonlinear analysis methods including stochastic resonance and manifold learning in this research area, the two studies which were isolated are combined to take advantage of each other’s strengths, and a unified research system is established, which is to utilize the multiscale and nonlinear analysis methods for enhancing machine condition features. Three parts are separately studied in depth as follows.The first part is the studies of the methods based on multiscale stochastic resonance analysis and their applications to weak condition feature enhancement of rotating machinery. The original envelope of machine measured signal is set as the treated object. The noise type, the weak feature and high-frequency characteristic of machinery fault in the envelope signal are analyzed. By introducing the method of stochastic resonance, and according to the influence of the noise at multiple scales on the system output, a new multiscale noise tuning stochastic resonance (MSTSR) method is proposed. The model of noise distribution can be tuned and optimized by the new method, and the weak high-frequency fault feature can be finally enhanced under the conditions of fixed noise level and type. A new index called weighted power spectrum kurtosis (WPSK) is constructed to evaluate the system output of MSTSR, by which the parameters in the method can be optimized without priori knowledge, and the optimized and enhanced result is finally produced adaptively. The multiscale noise tuning strategy is further improved, to meet the requirement of its application to multiple transient fault diagnosis. The multiple frequency components can be successively and separately enhanced without interferences from each other by the new improved method.The second part is the studies of the methods based on multiscale manifold learning analysis and their applications to inherent feature extraction of rotating machinery. The original measured signal is set as the treated object. Via combining phase space reconstruction (PSR) and continuous wavelet transform (CWT), a high-dimensional multiscale phase space containing signal non-stationary information is constructed. By applying the manifold learning on the high-dimensional phase space, a new time-scale manifold (TSM) signature is extracted, which can represent the inherent condition feature of supervised machine. The TSM ridge is extracted to directly demodulate the fault characteristic frequency on the time-scale plane, and a new exchange ridge demodulation method is proposed to improve the performance of in-band noise removal. In addition, the wavelet envelopes at multiple feature scales of measured signal are nonlinearly combined by using manifold learning method, thus a new feature called multiscale envelope manifold (MEM) is proposed. The new feature can represent the factual envelope structure of fault impulses. The applications of the proposed multiscale manifold learning demodulation methods to real bearing and gearbox fault diagnosis indicate their superiorities as comparing with traditional enveloping analysis methods.The third part is the studies of wayside acoustic signal processing methods based on multiscale and nonlinear analysis and their applications to railway train bearing wayside acoustic fault diagnosis. The bearing wayside acoustic signal is set as the treated object. A signal dynamic resampling method based on time-frequency distribution (TFD) instantaneous frequency estimation is first proposed to remove the Doppler effect contained in original signal. Then the MSTSR method is employed to enhance the weak fault characteristic frequency in the envelope signal after removing the Doppler effect. Two kinds of modulation phenomena related to the fault characteristic frequency possibly co-existing in the measured wayside acoustic signal are analyzed. A new method called variable-resolution ridge demodulation (VRRD) is proposed to extract the envelope of fault impulses from original signal directly. Then the dynamic resampling operator is conducted on the extracted envelope to directly rectify the characteristic frequency. Finally, through logarithmic transformation and piecewise linear detrending method, the possibly low-frequency modulation phenomenon is suppressed and the low-frequency nonlinear trend is removed. The results by simulation and experiment analysis confirm that the proposed methods can extract the fault characteristic frequency with improved signal-to-noise ratio and without the influence of Doppler effect from train bearing wayside acoustic signal.Moreover, the above research works were all validated by simulation and experiment analysis. Depending on the study objects and the methods being used, different simulation models are constructed for analysis in this paper. The analyzed experimental data include gearbox vibration signal with gear wearing fault, vibration or acoustic signals of different types of bearings with various localized faults, and railway train bearing wayside acoustic signals with different localized faults. The study of this paper indicates that, the multiscale and nonlinear analysis can enhance the machine condition features, which is of great significance for precision diagnosis of rotating machinery.