Aero-engine Vibration Fault Diagnosis Based On Hilbert-huang Transform

Vibration signals are the basic sources of information for rotating machinery's condition monitoring and fault diagnosis and these signals are usually non-linear, non-Gaussian non-stationary signals.The texture distributions and gray scale changes of the continous wavelet scalogram which extracted from continous wavelet transform can better reflect the non-stationary features of the faults. If we can extract these features and applied them to fault diagnosis, the early faults of the rotating machinery can be identified.By now, the proposed feature extraction methods are mainly 1-order gray moment vector based on wavelet coefficient matrix and texture feature of scalogram.However, the two kinds of scalogram feature extraction methods mentioned above do not analyse the non-linear characteristics of the faults, which only described the aspect of the second-order statistics of the images pixels and neglecting the images higher-order statistics. In response to these problems, this paper presents the use of Kernel Principal Component Analysis (KPCA) on the feature extraction of scalogram and at the same time use parameter adaptive support vector machine to classifiy the features extracted.Firstly, we introduce the basic principles of continuous wavelet transformation, the characteristics of wavelet basis function and its selection basis. Analyzed the failure mechanism, the spectral characteristics and scalogram image features of unbalance, misalignment, rab-impact and oil whirl fault.Secondly, we introduce the extraction methods of first gray moment vector and texture feature of wavelet scalogram. Study the basic principle of kernel method and several common kernel functions as well as discuss the principle method of principal component analysis (PCA). By combining kernel method with PCA; we raise the wavelet scalogram feature extraction method based on KPCA. We use ZT-3 multi-functions rotor simulation experimental system and aero-engine rotor fault experimental apparatus to obtain the samples of the four kinds of faults mentioned above, formed a sample set of 128 samples (with 32 samples of each type of fault). Transform these signals with continue wavelet transform to obtain the scalograms. Extracte the features above-mentioned fault samples and analyse the obtained data of these features.Thirdly, we study the principle and its advantages in learning classification of the support vector machine.In view of the current condition that there still no unified method to determine the model parameters of support vector machine, we study the effection of the kernel parameterσand penalty factor C of support vector machine. By using genetic algorithm, we optimize these two parameters, and constructe the parameter adaptive support vector machine model. Finally, by using this model, we classify the scalogram features extracted from the wavelet scalogram, compare the recognition rate and their effectiveness and study the different effections about different parameters. The result shows that scalogram features extracted by KPCA have a strong ability to identify faults and is helpful for intelligent diagnosis of rotor faults.