Study On Intelligent Diagnosis Method Of Three-Phase Full-bridge Controlled Rectifier Main Circuit Fault

With the rapid development of power electronics and continual emerging of novel power electronic devices, power electronic equipments have been increasingly applied to all aspects of industry and living. The requirement of its maintainability is also more and more important. Because there are many devices in the main circuits of high voltage, strong current systems, one will take much time and make many efforts to deal with it using normal fault diagnosis methods. Aiming at the own characteristics of fault, new technologies such as artificial intelligence, neural network, fuzzy algebra, wavelet analysis,etc have been successfully used to fault diagnosis of power electronics. Intellectualized automatic fault diagnosis method has the advantage of rapid analyzing and localizing the fault, cutting the stop time, increasing the efficiency and reducing the loss. This paper gives a new fault diagnosis method for power electronic main circuits: using voltage waveforms of power electronic circuit as fault information, analyzing the fault cases, and drawing out the fault characteristics as the inputs of neural network to realize fault diagnosis. The main works and conclusions in this paper are as following: (1) Three-phase full-bridge controlled SCR rectifier circuit is taken as an example to give fault waveforms. According to the practice, all sorts of the fault are analyzed. Then, the situations that can be classified are changed into fault information. The analysis shows that: a) The classifiable fault styles are mainly two types: one is fault occurs in single bridge arm, the other one is fault occurs in two bridge arms. Every type of fault can be classified into several classes. b) Under given triggering angle, the waveforms of the same type of fault only horizontally shift on the time axis and the shape of waveforms is not changed. c) Under different triggering angle, the shape of waveforms is changed. This will make drawing out fault characteristics more difficult. So, it is necessary to reduce the effect of triggering angle. (2) According to the multi-scale analysis theory, fault information's wavelet coefficient energy in four frequency band is changed into vector form as the input of neural network. a) Wavelet analysis is far superior to Fourier analysis in drawing out the signal features. Wavelet analysis has the characteristic of spatial locality, and the window wideness of the time and the frequency both can be adjusted, so it can analyze the details of a signal. Lacking of space locality in time domain, Fourier analysis is difficult to detect the spatial position and distribution of broken signal. b) According to Multi-scale analysis theory, signal is decomposed to nonintersecting frequency bands. The scope of each band is determined by the corresponding scale, as long as the mother wavelet and sampling frequency are given. c) According to energy analysis theory, wavelet coefficients are changed into vector form as the input of neural network. (3) BP neural network classifier is trained by the feature vector sets. The training process and result show: the structure is simple, training speed is fast and fault diagnosis result is exact.