The Research On Nonstationary Signal Based On Time-Varying Autoregressive Model And Its Application In Fault Diagnosis

The mechanical fault diagnosis is vital significant in the modern production. It is an important method of fault diagnosis that withdrawing the information and the parameter from the vibration signals which reflect device vibration state. However the fault occurrence and the development, the change of equipment operating condition as well as equipment owning non-linearity can lead the dynamic signal of mechanical device to display the nonstationary characteristic. The traditional signal processing methods suppose the signal stability or part stability, so it must influence the accuracy of the analysis results. It will be helpful to enhance the signal-to-noise ratio and fault diagnosis quality carrying on accurately modeling to the nonstationary signal, extracting useful information and suppressing the disturbance of other ingredients.This article that bases on literature investigation and material research has analyzed the present situation of the existing vibration signal analysis technique and the mechanical fault diagnosis technique. What is more, it has proposed a parameterized time-frequency analysis method which named Time-Varying Autoregressive (TVAR). This method has the unexampled superiority to the traditional non-parameterized time- frequency methods. The fault diagnosis simplicity and the accuracy has greatly enhanced by using extracting time-frequency characteristic from time-frequency spectrums which obtain from Time-Varying Autoregressive method and going on the automatic diagnosis through BP neural network.In view of the vibration unstability of complex equipment system, this article proposed a new parameterized nonstationary signal processing method, named Time-Varying Autoregressive (TVAR). This method overcomes the disadvantages of traditional time-frequency analysis method effectively such as low resolution, existing cross term disturbance and so on. Besides, it also enhances the accuracy and the reliability of fault diagnosis obviously. Different type signals select the different basis functions. At present, how to choose appropriate basis function according to the signal own characteristic still has no effective method and criterion. Therefore, this article studies the influence of the different basis functions to the analysis results. Finally, it determines the application range of different basis functions. Moreover, this article also introduces several kinds of model order determination algorithm because the model order of TVAR model affects the frequency estimate precision directly.Today, the characteristic extraction is a bottleneck problem of mechanical fault diagnosis. It relates the accuracy of fault diagnosis and the reliability of early forecast directly. Therefore, it is an essential important link in revolving mechanical fault diagnosis that withdrawing the character information from the gathering primary data and guaranteeing the fault diagnosis to run effectively and accurately with the aid of signal processing and other methods. This article introduces the fault characteristic extraction methods based on the time-frequency analysis and time-frequency spectrum. And finally it has determined the characteristic extraction method which based on the energy distribution of time-frequency spectrum.The experiments has proved that time-Varying Autoregressive algorithm can set up the parameterized modeling for nonstationary signals accurately and gain the high-resolution time-frequency spectrums. Meanwhile, using the fault characteristic extraction technique based on energy characteristic from the time-frequency spectrums can extract the fault characteristic effectively, and provide the advantageous technical support for the accuracy of neural network fault automatic diagnosis. Furthermore, the automatic diagnosis through BP neural network can avoid the man-made diagnose mistakes.