Fault diagnosis with general parameter radial basis function neural network

Summary form only given. Fault diagnosis is one of the important and complex problems of modern control theory and practice. The choice of approach depends on plant characteristics, production conditions, technological variable information available and economical limitations. These factors are the basis for optimal plant description form determination. The general parameter (GP) approach, which is characterized by high convergence rate at model learning stage and simple decision-making procedure at diagnosis stage, is presented. The nonlinear time series model reconstruction problem may be solved by a neural net approach. There is a trade-off between approximation ability, structure complexity and convergence rate. A neural net structure formation and learning method which has faster training features is proposed based on GP methodology. Fast convergence is achieved by initial reduction of net parameter space dimensionality. The latter is increased then during training, if necessary, to improve approximation ability. Improved structure radial basis function networks is developed. Simulations are performed to confirm the feasibility of this net. The method is experimentally applied to a machine tool breakage detection. It has low computational time.