Research On The Method Of Feature Extraction And Pattern Recognition For Bearings Fault Of Rotating Machinery

The fault diagnosis technology for rotating machinery is developing rapidly from theory research to practical application in parallel with the new achievement of modern science and technology, and becomes a new interdiscipline composing of mathematics, computer, signal processing and artificial intelligence, etc. Rolling bearing is an universal part of rotating machinery. Therefore, fault monitoring and diagnosing of rolling bearing is a hot research in the advanced mechanical field. The random vibration signals of bearings and components are employed in monitoring and diagnosing, which is the common used method in the study of mechanical fault monitoring and diagnosis.The process of machinery fault diagnosis includes the acquisition of information and feature extraction and pattern recognition of which feature extraction and pattern recognition are the priority. In terms of the analysis of fault vibration signal feature of rolling bearings, the general fractal dimension of multi-fractal theory, AR model, wavelet analysis and principal component analysis methods are employed in the dissertation and the feature of the rolling bearings is achieved. The research shows that the fault feature of rolling bearings is extracted effectively. A method of AR model and radial basis function (RBF) networks are combined and applied to the rolling bearing fault diagnosis. Wavelet packet analysis and RBF networks are also combined and applied to the rolling bearing pattern recognition. The analysis results show that the two fault diagnosis systems can classify working condition and fault patterns accurately. Support vector machine (SVM) are introduced into rolling bearings intelligent fault diagnosing due to the fact that it is hard to obtain enough fault samples in practice. A novel method of principle component analysis (PCA) and SVM are combined and applied to the rolling bearing fault diagnosis. The analysis results from rolling bearing vibration signals show that the fault diagnosis method based on PCA and SVM are effective. The research is carried out in the following aspects:1. The basis theory and applications of fractal method are discussed and the general dimensions algorithm of multi-fractals theory is underlined. The calculation and simulation of analog signals general dimensions verify the algorithm, pave the way for the fractal fault diagnosis of rolling bearings. The results show that any general dimension is able to demonstrate the feature of a certain non-complex signal such as sine wave, whereas multi-fractals dimensions should be employed to describe a complex signal. It is necessary to illustrate the set or sample length of general fractal dimensions in calculation.2. Multi-fractals theory and general fractal dimensions methods are employed in the fault diagnosis for rotating machinery, and the testing system and analysis software are presented. The general fractal dimensions sample library is established for the typical signals in different conditions, the general fractal dimensions spectrum of typical samples is obtained. The results show that the rolling bearings vibration signals have similar fractal dimensions in the same working condition, whereas, the fractal dimensions are different when the working conditions are different, the difference is distinct. It is accurate and feasible to evaluate the bearings fault with the method of general dimensions.3. General dimension correlation coefficient method and general dimension series single value approach method are employed to achieve the fault pattern recognition of rolling bearings. The theory and experiment analysis show that the fault pattern of single fault type is obtained by compare the calculated general fractal spectrum with the general fractal spectrum in sample library. However, the accurate fault types can not be recognized with above method when the signals are complex and coupling. It is necessary to combine the general dimension correlation coefficient method and general dimension series single value approach method to achieve the recognition. The general dimension correlation coefficient method is a preferable method in fault pattern recognition for the few samples and high recognition rate. The working conditions and the fault pattern recognition of rolling bearings are achieved clearly by different general fractal dimensions without disassembly. The simple and feasible method is a new approach to the rolling bearings fault diagnosis. 4. The modeling, order and feature extraction of the AR model is discussed, and the intelligent diagnosis model, learning rules and the algorithm of the RBF neural networks is studied. A method of AR model and RBF neural networks are combined and applied to the rolling bearing fault diagnosis, the detailed algorithm and analysis software are presented. The key issue of the fault diagnosis is to extract the feature which indicates the fault information rapidly and effectively before the diagnosis system is established. The experiment and analysis results show that if the training scale is correctly chosen and feature parameter of the AR model is evaluated reasonable, the fault pattern of rolling bearings is recognized accurately.5. The applications and limitations of conventional FFT and STFT in signal processing are briefly discussed. The wavelet analysis method is introduced into rolling bearings intelligent fault diagnosis, the wavelet packet denoising and the eigenvector extracting of wavelet packet is presented. A method of wavelet analysis and RBF neural networks are combined and applied to the rolling bearing fault diagnosis, the detailed algorithm and analysis software are presented. The experiment and analysis results show that the wavelet packet denoising and decomposing method is effective. The fault feature is presented by the each frequency component energy directly and the original frequency feature is maintained, the eigenvector of wavelet packet denosing and decomposing is the input of RBF neural networks. The results show that the proposed method can recognize the fault pattern accurately.6. SVM are introduced into rolling bearings intelligent fault diagnosis due to the fact that it is hard to obtain enough fault samples in practice and the perfect performance of SVM. The two-class classifier performance of SVM is discussed under the different conditions such as different sample quantities, different types of kernel function and different parameters in the same kernel function. The experiment and analysis results show that SVM have perfect classified performance in only limited training samples and the diagnosis precision is less dependent on the kernel function and the parameter, which is suitable in the engineering applications. This dissertation offers a comparison between two classification algorithms, RBF networks and SVM for cases where only limited training samples are available for diagnosis. The results show that SVM have better performance than RBF networks both in training speed and recognition rate.7. A novel method of PCA and SVM are combined and applied to the rolling bearing fault diagnosis and the feature choosing and feature extracting are discussed. The dimensions of AR model parameters of time-sequence analysis is reduced by PCA and the vectors are inputted into the SVM classifier. The experiment and analysis results show that PCA method can effectively reduces the dimensions of eigenvector, so reduces the calculating complex degree of the fault classifier. The three-class classifier performance of SVM is discussed where only limited training samples are available. The analysis results show that diagnosis precision is not varied basically before/after the dimension reducing, the SVM method have perfect classified performance in only limited training samples and the diagnosis precision is less dependent on the kernel function and the parameter. The analysis results show that the fault diagnosis method based on PCA and SVM can extract rolling bearing fault features effectively and classify working condition and fault pattern accurately.