EMD And BP Neural Network Based Fault Diagnosis And Its Application

In this thesis some Time-Frequency analyzing method are reviewed first, analyzing their advantages and disadvantages. Then a new analyzing nonlinear and non-stationary signal method is introduced. And a comparison between it and the other Time-Frequency analyzing method is done, pointing out the new method's advantages. On this basis, using it associated with BP neural network in fault diagnosis is investigated thoroughly.Most of the former Time-Frequency analyzing methods are derived from the Fourier analysis. So they more or less have the problem exiting in Fourier analysis, for example: the discontinuity expressed with harmonic components, being integral mean of a time interval; and the problems derived from the conventional definition of frequency, that is, the contradiction between the time and frequency resolution. Improving the performance of some parameter is at the cost of sacrificing the other one.A powerful tool for analyzing nonlinear and non-stationary signal which is called EMD (empirical mode decomposition) method, is introduced. Different from the traditional method in doing integral transformation to signal, it decompose signal into several IMFs (intrinsic mode function), which contain and extrude the local characteristic of signal. So the characteristic information of the original signal can be well held by analyzing the IMFs. The experimentation testifies that the EMD method applies to linear and stationary signal. At this circumstance, the decomposed IMFs are the frequency components. In fact, EMD is a method which extract characteristic from signal. So in most circumstances, it is combined with other method to analyse signal.The essence of using neural network in fault diagnosis is a matter of pattern recognition. The emphasis and difficulty lie in how to extract characteristic of signal efficiently, that is, the problem of seeking classification criteria. The appearance of EMD method provides a new way to solve this problem. Fault diagnosis experimentation is carried out. In the experimentation the IMFs decomposed fromsignal are input into a designed BP neural network to train the network. Through this experimentation it is verified that the method of EMD in conjunction with BP neural network can not only improve the distinguishing accuracy, but also reduce the learning time of BP neural network greatly. What's more, the study covers how to decrease the number of IMFs and simplify the structure of BP neural network. A experiment proves its feasibility.At the end of the thesis, the application prospect of EMD in fault diagnosis is revealed. The disadvantages and development direction of EMD are pointed out.