Off-line identification of nonlinear, dynamic systems using a neuro-fuzzy modelling technique

This paper presents a methodology for generating training data for use in identifying a type of neuro-fuzzy model: a fuzzy relational model. Issues associated with identifying accurate neuro-fuzzy models of nonlinear dynamic systems are discussed and the importance of finding a suitable method for generating the input–output data used to estimate the parameters of the model is explained. Different ways of generating the training data are compared and a new method of directly generating the training data is proposed. Two excitation signals are used to generate the data. The first consists of a series of step changes between values at the apexes of the fuzzy sets describing the input variables. The second is a chirp signal that excites a range of frequencies over the bandwidth of the system to be modelled. Results obtained from a simulated water-level control system are used to demonstrate that the proposed methodology can successfully identify a satisfactory fuzzy relational model of the system, and show that the performance of the resulting model is very sensitive to the type of test signal used to validate it.