Aggregative Learning Method and Its Applications to Fault Detection

Abstract

A novel approach for mathematical model reconstruction, called Aggregative Learning Method (ALM), is proposed Realizations of the developed approach for different model types are discussed. Among them are regressive and autoregressive linear and nonlinear models, neural network models, neuro-fuzzy models. The abilities of the method to solve a wide class of problems, such as parameter identification, structure identification, prediction, automatic fuzzy rule set generation, are studied. It is shown that the method has effective training features and provides information about model quality achieved during the learning stage. Improved structure backpropagation and radial basis function networks are developed. A fuzzy procedure is designed for quasi-optimal network structure determination based on learning stage information. Simulations are performed to confirm the feasibility of proposed procedures for practical applications. Next, these are practically applied to nonlinear time series identification, prediction, and fault detection problems resulting in higher mapping accuracy, faster learning speed, higher fault sensitivity, and decision reliability in compare to conventional approaches.