Study On Fault Diagnosis For Systems Based On Alpha-PNN

With production process becoming more and more complicated, how to improve the dependability and security of control systems has already aroused great concern from people. The application of artificial neural network techniques in fault diagnosis has become an important research area in recent years. Probabilistic neural network (PNN) has been extensively adopted in fault diagnosis due to its features of simple network learning process and quick training. It has become an important method for fault diagnosis. But, probabilistic neural network also exist some problems, such as how to get the value of smooth parameter for better classification and how to choose the number of neurons in the hidden layers.This paper connects with related theory about alpha stable distribution and probabilistic neural network according to the key problems of probabilistic neural network need to be solved in fault diagnosis. Two typical fault detection and diagnosis methods are studied. One is robust fault detection of system based on parameters estimation of symmetric alpha-stable distributions while the other is fault diagnosis of system based on alpha-PNN. The main contents and conclusions are listed as follows:Firstly, a new fault detection method based on parameters estimation of symmetric alpha-stable distributions is proposed for the nonlinear systems which are difficult to model. Firstly, the sampling data of output are regarded as time series. The output series is predicted, and the prediction error signal with obvious abrupt impulses is obtained. Then the value of parameter alpha can be estimated through the method of parameter estimation of symmetric alpha-stable distributions. Therefore, the curse of parameter alpha is established. It is explicit to detect system fault on the basis of this curse.Secondly, in order to overcome traditional PNN's assumption that input data are independent and identically distributed, a new modified probabilistic neural network named alpha-stable distributions basis function probabilistic neural network is presented, and a new fault diagnosis menthod is also proposed. The activation functions of network's hidden neurons adopt probability density function of symmetric alpha-stable distributions. Compared with routine gauss distribution function, it has better variability and tractility, so hidden neurons have high adaptability in terms of function approximation, meanwhile, it overcomes PNN's assumption that input data are independent and identically distributed, and it also improve neural network's approximation ability of partial pulse burst. The simulation result indicates that this algorithm also can improve the recognizing effects and have low false positive ratio than PNN with the assumption of colored noise.Finally, the paper gives a brief summary of the job as well as the future expectations.