Research On Mechanical Fault Feature Extraction Method Based On Non-gaussian And Non-stationary Signal Processing Theory

Mechanical fault diagnostics is an integrated technical subject based on mechanics. The key of mechanical fault diagnosis is how to extract the fault feature from mechanical fault vibration signals which are non-gaussian and non-stationary signals. Recently, in order to meet the need of early detection and diagnosis on mechanical faults, non-gaussian and non-stationary signal processing methods attract extensive attention in the area of mechanical fault diagnosis. How to creatively apply non-gaussian and non-stationary signal processing theory to solve the problems about signal denoising and feature extraction is one of the important research issues in the area of mechanical fault diagnosis. Based on the above request, this dissertation does some researches on non-gaussian and non-stationary signal processing methods suitable for mechanical fault feature extraction. The main research work by the authors is introduced briefly as follows:Firstly, vibration signal demodulation analysis and frequency spectrum Zoom-FFT technologies based on STFT are studied. From the point of view of signal filtering, the principle of vibration signal demodulation method based on STFT is described and some factors affecting the demodulating performance are discussed. Through theoretical analysis, it is pointed out that the vibration signal demodulation method based on STFT is a special demodulation method based on complex analytic band-pass filtering and Hilbert transform. It is theoretically proven that the signal, which has the same period with the modulated signal, can be restored by demodulation analysis if some frequency elements of the modulated signal are located in pass range of a complex analytic band-pass filter. Based on the above analysis, the novel practical STFT based adaptive vibration signal demodulation algorithm is proposed. About the problem of weak periodic fault feature extraction in complicated background noise environment, a new detection method based on the combination of singular value decomposition denoising and STFT demodulation technique is put forward. In addition, an improved STFT based frequency spectrum Zoom-FFT technology is proposed.Secondly, vibration signal processing technologies based on filterbank theory are studied. For overcoming the deficiency of current discrete wavelet and wavelet packet transform based vibration signal decomposition method, the decomposition method based on quadrature mirror filterbank (QMF) is first proposed and a two channel QMF with linear phase is constructed. Compared with the wavelet filters, constructed QMF filters with the same filter order as the wavelet filters have better filtering performance and the filter coefficients can be obtained more easily. In order to solve the problem of the frequency band derangement in the normal two channel filter band decomposition algorithm, an improved QMF decomposition algorithm is proposed. Based on the improved algorithm, a new vibration signal demodulation method applied to automatic detection of early mechanical fault and a new frequency spectrum zoom method are introduced. Given the deficiency of two channel pyramid decomposition, a three channel signal pyramid decomposition scheme is put forward as the supplement of two channel pyramid decomposition. In addition, referring to the stationary wavelet packet decomposition algorithm, a new vibration signal stationary filterbank decomposition method based on QMF is proposed. The analysis results on simulated and measured data show that the filtering performance of the stationary filterbank decomposition method based on QMF is better than that of the stationary wavelet packet decomposition method.Thirdly, weak impulsive response signal feature extraction method based on continuous wavelet filtering is studied. From theoretical and simulated experimental studies, it is shown that the normal wavelet series base function and the normal time-scale analysis method are not suitable for feature extraction of weak impulsive response signal. Based on the uniform form of continuous wavelets suitable for feature extraction of weak impulsive signal, a new frequency domain compact support wavelet whose spectrum center frequency and frequency spectrum window width can be easily adjusted is constructed. The simulated data show that this wavelet filter can effectively enhance the impulsive signal feature when the selected wavelet parameters are reasonable. On how to rapidly design optimal frequency domain compact support wavelet suitable for weak impulsive signal feature extraction, the adaptive design method based on kurtosis as optimal objective function and using genetic algorithm to optimize parameters is proposed. In addition, the adaptive wavelet preprocessing method is applied to second order cyclostationary signal demodulation analysis of the weak impulsive modulation signal. The improved spectrum correlation density demodulation analysis method based on the preprocessing of optimal frequency domain compact support wavelet is first proposed. The analysis results on simulated and measured rolling element bearing data show that the improved method not only can effectively demodulate the fault feature frequency of weak periodic fault impulsive signal, but also can greatly reduce the computing burden and improve the practicability of second order cyclostationary signal demodulation method.Fourthly, empirical mode decomposition (EMD) method on vibration signal is studied. In order to solve the problem of endpoint effects in empirical mode decomposition process, a new endpoint extreme value extrapolation method based on the comparison of waveforms is proposed. The analysis results on simulated and measured rotor vibration data show that, for EMD decomposition of the regular signal, using the new method can effectively reduce the negative influences of endpoint envelope error to EMD decomposition result and obtain true IMF components. For possible appearance of mode mixture in EMD decomposition, it is pointed out that various certain energy level irregular stationary and non-stationary additional noises are the source of mode mixture. Based on the above conclusion and understanding of the EMD method in nature, a novel adaptive filtering based mode demixing method is proposed. The analysis results on simulated and measured gear vibration data show that this new mode demixing method can effectively reduce the influence of mode mixture on decomposition result.Finally, Blind vibration signal processing method and its application is studied. By analyzing implement theory of frequency domain multi-channel blind deconvolution algorithm, a multi-channel vibration signal blind feature extraction method based on the combination of stationary filterbank decomposition filtering and ICA algorithm is proposed. This new blind feature extraction method is applied to extract turbocharger shaft rotating frequency feature of locomotive diesel turbocharger vibration signal in serious background noise environment. The result shows that using this method can effectively extract shaft rotating frequency feature of the weak turbocharger vibration signal. Given the deficiency of normal single channel vibration signal blind deconvolution algorithm—minimum entropy blind deconvolution algorithm, a new single channel vibration signal blind deconvolution algorithm based on ICA is proposed. The results on simulated and measured vibration data show that the new algorithm has more rapid velocity of convergence, better robustness and can more effectively extracts weak impulsive response signal feature.