Research On Feature Extraction And Representation With Vibration Signal For Rotary Machinery Condition Monitoring And Fault Diagnosis

Fault feature extraction and representation is the most crucial problem for the reliability and accuracy in the mechanical condition monitoring and fault diagnosis. This dissertation explores the applications of the theories with absolute value of autocorrelation, wavelet fractal analysis, independent component analysis and blind source separation in the feature extraction, representation and classification of vibration signal for the rotary machinery, such as bearing and gear.In chapter 1, at first, the significances and content of mechanical condition monitoring and fault diagnosis are pointed out, the current detection scheme and technical means are reviewed for the condition monitoring of bearing and gear, then the applications of the time domain statistical analysis, time-frequency representation, independent component analysis and blind source separation are reviewed as well, lastly, the contents, the focuses and the innovations of this dissertation are pointed out.Detection of signal transients based on the absolute value of autocorrelation and the representation of enforced fault feature in the polar coordinates are proposed in chapter 2. The new method is done in four steps:first, calculates the time average function of the original signal; second calculates the autocorrelations of the time average function; third, calculates the spectrum of the autocorrelations by fast Fourier transform; fourth, the enforced fault feature is represented in the polar coordinates corresponding the possible periods of the transients. The experimental verification shows the effectiveness of the new proposed method for bearing condition monitoring.The applications of the fractal analysis based on wavelet transform to the classification of bearing's vibration signals are studied in chapter 3. Firstly the acceleration signals are decomposed to detailed signals at different wavelet scales by using the discrete wavelet transform. The variances of detailed signals are calculated and then fractal dimensions of the acceleration signals are estimated from the slop of the variance progression. The fractal dimensions are significantly different among the different working conditions of the bearings and showes a high reproducibility. The results suggest that the wavelet-based fractal analysis is effective for classifying the working conditions of bearing.The basic theory of independent component analysis and its application to feature extraction from one-dimension vibration signal are studied in chapter 4. Firstly, the ICA general model and the dissimilarity between statistical independence and uncorrelated property are reviewed, then the contrast functions:kurtosis, negentropy and mutual information for measures of nongaussianity are presented, also is the centering and whitening for preprocessing for ICA. Secondly, the FastICA algorithm is studied. Finally, a new feature called ICA filtered correlation feature is quantitatively calculated by the transformed coefficients. The new feature has the clear class property and can be applied for signal classification.Blind source separation for convolutive mixtures and its application to machine vibrations are studied in chapter 5. First reviews the general concept of blind source separation, especially the theory for convolutive mixtures, the model of convolutive mixture and two deconvolution structures:Recursive and Direct structures, then presents a BSS algorithm for convolutive mixtures based on RCTE threshold control criteria, last the simulation testing points out good performance for simulated mixtures.Chapter 6 gives the conclusions and the prospect about this study.