Investigation On Mechanical Fault Feature Extraction And Classification Based On Independent Component Analysis

With the modernization of machinery manufacture, condition monitoring and fault diagnosis become more and more important for machine equipments. Due to the structural complexity and working integration of modern equipment, the traditional signal processing methods can not accurately extract fault features from the observed mixed-signals. However, BSS technique can recover the vibration source signals through the statistical properties of observed signal, then select one or a few vibration source signals for further analysis. The BSS method has become a research focus in many study areas. This dissertation has done some research on the blind source separation method and its application in mechanical failure diagnosis filed. The primary work is as following:1. Based on the analysis about the decomposition ability and the mode aliasing of EMD, a new method called EMD-ICA is proposed to overcome the application difficulty of ICA on one-dimensional mechanical fault signal. And the definitions of EMD-ICA process feature and EMD-ICA estimated sub-band feature are expressed, and the feature information indicated fault component obviously is used to the inputs of the probabilistic neural network, so the different mechanical fault identification and classification are accomplished. The advantage of this method includes the following several parts:the undertermined problem of ICA in mechanical fault diagnosis is solved; the analytical deviation caused by the weak distinguishes and redundancy is reduced; the computational complexity of neural network is greatly decreased; the accuracy of fault recognition is improved.2. In order to extract the weak information from one-dimensional incipient mechanical fault signal, a new method called phase space reconstruction ICA and the contribution coefficient of kurtosis is put forward based on the intensive study about the phase space theory and ICA temporal structure theory. The phase space embodied incipient mechanical fault is reconstructed, the phase space components are selected as the inputs characters of ICA algorithm, and estimated components involved weak fault feature is reconstructed, then the conception of the contribution coefficient of kurtosis is proposed to extract the position and period of incipient fault from reconstructed signal. According to an example analysis, it can be concluded that this new method can well extract weak impulse information in the situation of weak fault feature information and low signal-to-noise ratio based on phase space reconstruction ICA and the contribution coefficient of kurtosis.3. Two improved time-frequency SDICA methods is expressed on the basis of view and study of the time-frequency ICA application in mechanical fault diagnosis. One is HHT SDICA method based on wavelet packets decomposition; it can diminish the redundancy information of different fault time-frequency information, so the projection coefficients of different fault time-frequency information are used as the neural network inputs, then the accurate classification of different fault is received. The other is SPWV SDICA method based on IIR filter, it can eliminate the cross-terms of SPWV in the SDICA application on SPWV and improve the accuracy of SPWV SDICA separation. At the same time, the extensible research of SDICA is explored based on time-frequency ICA, a SDICA with selection criteria is advanced to the BSS of statistical correlation images.