In-depth Extraction Of Nonstationary Feature For Vibration Signal And Development Of Remote Diagnosis Service System

The vibration signals of mechanical equipment often represent nonstationarities due to occurrence and variance of fault, influence of nonnormal working condition and inherent nonlinearity of equipment. Nonstationary feature representation and extraction is the most crucial and difficult problem for reliability and accuracy in mechanical fault diagnosis. Though the time-frequency representation is a powerful tool in the study of nonstationary and nonlinear signal, the traditional time-frequency analysis theories and methods are not always suitable to describe the nonstationarities with the increasing complex of new fault,. In this dissertation, based upon a summary of the present worldwide research status, the nonstationary feature of machine vibration signal is defined and the significance of its deep-seated extraction is analyzed. It is indicated that the nonstationary feature of machine vibration signal mainly focus in some main physical quantities, such as the instantaneous frequency, phase ,amplitude, energy , modulated information etc., all of which can be extracted only with further processing the analyzed result from traditional time-frequency distribution, and then can be used for diagnosis. Taking bearing and gearbox being as the researched objects, some practical techniques for deeply extracting nonstationary feature based on traditional time-frequency analysis, is proposed and developed.In the aspects of the nonstationary feature in-depth extraction, the main contributions of this dissertation as follows:(1) Nonstationary feature extraction techniques based on Wavelet Transform (WT) are studied. The theory and engineering significance of WT are introduced. The selection problem of wavelet basis and transforming scale is analyzed in detail. With complex analyzed wavelet transform, the instantaneous amplitude and phase can be computed, and then the instantaneous frequency can be extracted with the wavelet ridges method. The feature representations of wavelet scale-energy and time-energy correlation are proposed, and are used to represent the fault features of bearing roll, inner race and outer race successfully.(2) Nonstationary feature representation and extraction method based upon the Wavelet Packet Decomposition (WPD) and demodulation techniques are studied. With WPD, the vibration signal can be denoised, and some component in special frequency band is selected to reconstruct. With the decomposition and reconstruction of wavelet packet, the monocomponents with fault feature in different frequency band would be captured and separated out. Based upon this, Teager energy operator demodulation method and improved Hilbert instantaneous frequency estimation method are introduced into WPD to be a syncretic technique fitting for extracting noncomponent instantaneous frequency and modulation information from multicomponent complicated signals. This syncretic method is applied into the gearbox fault diagnosis and achieves good effects.(3) Nonstationary feature extraction based on the joint Time Frequency (TF) analysis is studied. The conceptions of linear and nonlinear TF transformation, their properties and cross-term interference problem are summarized and analyzed systematically. On the basis of traditional self-adaptive TF analysis, three new techniques for fault signal processing are proposed. One is the improved Chirplet self-adaptive TF representation, in which an additional parameter referred to as curvature parameter is introduced in the traditional Chirplet elementary function to match the time varying linear or non-linear components. The performance comparison between the modified version of adaptive TF representation and the other TF representation, verifies that the modified version has the high TF resolution and no cross-term interference; the other is the TF decomposition based on curved surface fitting, which has been applied successfully into nonstationary signal denoising, another is the transient impulse signal extraction based on self-adaptive TF decomposition, which uses the vibration impulse signal model to approximate the signal filtered out in TF plane, in order to determine the parameters of impulse signal model. On the basis of self-adaptive kernel function TF representation, the precise IF extraction method using TF ridges is proposed. In this method, by segmenting in time domain with the way of adding time windows, the whole domain is cut into many sub-segments and the whole Redial Gaussian Kernel time-frequency Distribute(RGKD) can be obtained by concatenating the sub-segment RGKD, and then, the whole RGKD is regarded as an image, the ridges of the image are extracted using digital image processing method including smoothing operation, 2-D Laplacian arithmetic operator and thinning, finally, the principle of detecting the lines of Hough transform is used. Based on the instantaneous frequency extracted result, the TF filtering method based on instantaneous frequency is developed, by which the special monocomponent is separated from the original signal.(4) Nonstationary feature extraction base on Empirical Mode Decomposition (EMD) is studied. Based on the basal principle of empirical mode decomposition (EMD), three existing problems from EMD, that is, end effects, the select of sifting stop guide and mode mixing, are analyzed. For these limits of conventional EMD, some improved methods are proposed for the precision and correctness of EMD, combining the problems occurring in practical applications. Two novel solutions to mode mixing problems are presented in detail. One is expending best wavelet packet decomposition (WPD) into EMD, that is, applying WPD based on Shannon entropy into the selected (IMF) in which the mixing mode exists. The second novel solution to the problem of mode mixing is the masking signal method. The experiment results verify that these two methods both are benefit to the improvement of EMD decomposition ability.Our final intention is applying all these nonstationary feature extraction techniques to solve the practical engineering problems, transforming the science technique to productivity. A number of practical signal analysis examples indicate that synthetically applying above feature extraction methods, the exact diagnosis result can be attained. As the carrier of these extraction techniques, the remote measurement and diagnosis service system proposed in the dissertation realizes the amalgamation of E-commerce, information network and industry intranet. Its overall architecture are researched and designed, and some key techniques including virtual instrument,DataSocket remote data transferring,neural network study and diagnosis,network security technique etc. are introduced in detail.The methods of nonstationary signal feature extraction developed in this dissertation take on the important reference value for mechanical equipment fault diagnosis. The remote measurement and diagnosis service system in E-commerce environment can cut the time of collecting fault, improve the efficiency of diagnosis system, and favor accumulating data and sharing resources. All the studies are practical and realistic, and full of innovation. The research productions are beneficial for theory study and development practice in the nearly future.