Study On Nonstationary Feature Extraction And Intelligent Diagnosis Based On Parametric Model

As the complicated and automatic development of machine, there is more and more effect of intelligent diagnosis in modern manufacturing industry. In order to get higher identification ratio, the fault diagnosis technique based on the combination of signal processing and neural network model has become the host research field. Actually, the vibration signals of mechanical equipments have nonstationary charac- teristics under complex working condition. Therefore, it is very important to develop the research on features extraction of nonstationary signals and automatic fault identification. Time Varying Aoturegressive model (TVAR) and Generalized autoregressive conditional heteroskedaticity model (Garch) were studied in this paper for analyzing nonstationary signals. And based on features of parametric model, the HMM and SVM were studied for classifying various faults of rolling bearing automatically. Meanwhile, the main contents as follows:First, aiming at the parametric model of nonstationary signals, the TVAR model was studied. The time varying parameters and spectrum estimation were solved by adaptive arithmetic and basis function arithmetic respectively. And the AIC rule was adopted to determine model's order. Through analyzing the simulation signal, it is proved that time varying spectrum had high resolution and not cross term. However, the TVAR spectrum based on basis function arithmetic, was able to track signals which changed fast, had little influence to noise and could make up for the short- coming of adaptive arithmetic model which was sensitive to noise. Meanwhile, aiming at the problem that TVAR model couldn't track the mutational signals, Garch model and its analyzing steps were studied in the paper. Then, through analyzing the simulation signal, it was proved that the mutational characteristic or impulsion was reflected in the Conditional Heteroskedaticity of Garch model accurately. Then, the diagnosis theory and method of SVM and HMM model were studied.First, the problem of multi-class classification was solved by"one vs all"method; Besides, SVM model frame and the fault diagnosis steps were discussed. Meanwhile, the training arithmetic and fault diagnosis steps of HMM and features quantization of signals with Lloyd arithmetic were studied. Then, fault diagnosis of SVM and HMM was realized based on Matlab. Through practical example of rolling bearing, it was proved SVM model had excellent learning capability and good classification capacity under small sample. So the problems of neural network that much sample and local maximum could be solved by SVM model. In additional, HMM had advantages that high judgement precision and good self-study ability. Especially, aiming at the characteristic of mechanical vibration signals that strong nonstationary and bad reproducibility, it had high identification ratio.At last, on the above research foundation, the software for parametric model and intelligent diagnosis of rolling bearing signals was exploited based on labview. The functions that time frequency analysis of TVAR and intelligent diagnosis based on SVM and HMM were realized on windows XP platform. Therefore, the research results could facilitate to develop the features extraction and intelligent diagnosis of nonstationary signals.