Application Of Fault Diagnosis Methods Based On Local -Wave Method And Blind Source Separation

Mechinery diagnosis is a branch of mechanology. Its essence is the pattern recognition of machine operating condition. Its key issues are the feature extraction and classification for fault signals, hi recent years, to meet the needs of early detection and accurate diagnosis of mechanical faults, non-stationary and non-Gaussian signal processing techniques attracted more and more attention in mechanical fault feature extraction, In this dissertation, based on " Research on Local Wave method and its Engineering Application" (National nature science fund project. No:50475155 ), combined with the pattern recognition and machine learning techniques, the problems of feature extraction and fault diagnosis are addressed using Local Wave method, Wigner higher-order time-frequency representation and blind source separation of non-stationary, non-Gaussian vibration signals. The main contents as follows:1. Local-wave method is studied and applied to mechanical vibration non-stationary signal. By comparing with wavelet and several time-frequency methods, the Local-wave method can be proved to more effective than others. The experiment result shows that Local-wave time-frequency analysis can clearly explain time varying characteristics of different fault modes. But this time-frequency method is a two-dimensional signal representation. It arises the dimensionality problem. To describe the signal with as few variables as possible, the geometric moments and margins were used as the features of the time-frequency distribution. Then, combined with the artificial neural network, the fault diagnosis method is proposed based on geometric moments and marginal densities of Local-wave.2. The effectiveness of a multi-component neural-network architecture based on Local wave for the time series prediction of non-stationary, nonlinear dynamic systems has been investigated. The simulated experiment for sunspots' benchmark suggests that the multi-component architecture outperforms the corresponding single-scale architectures. Then, an observer bank of autoregressive time series models based on multi-component neural-network architecture is used for model diagnosis of rotor fault vibration signals. These models can be used as one step ahead predictors allowing comparison of signals for the purposes of fault detection and diagnosis. From the prediction error, features can be extracted and used to determine the machine's condition. Vibration data from a rotor placed under different fault conditions were used for training and testing models. The experiment results indicate that this approach could be used to diagnose fault conditions.3. Higher order time-frequency distributions and their applications to feature extraction of machine vibration signals are studied. The detection of these impulses can be useful for fault diagnosis purposes in condition monitoring environment. The features can be effectivelyextracted for these non-stationary, non-Gaussian vibration signals by the Wigner higher order moment spectrum. The Local wave decomposition method is used to suppress the cross-term for multiple signals in higher order time-frequency distributions and the simulated results are satisfactory. The vibration signals measured from diesel engine in the stage of deflagrate are analyzed with Wigner higher order moment spectrum. Experimental results indicate that this method has good potential in mechanical fault feature extraction and can be used to extract the valuable quantifiable information about the time-frequency features of impacts.4. The blind source separation fault diagnosis method based on Local wave time-frequency images is developed. ICA has been mainly used on the problem of blind signal separation. ICA is a feature extraction technique which can be considered a generalization of principal component analysis (PCA). ICA can be used to gain the local features which give better image representation. Because the Local wave time-frequency image can reflect the character of different fault conditions, blind source separation as a new signal pr...