Generating training data for identifying neurofuzzy models of non-linear dynamic systems

This paper presents a methodology for generating data for training a fuzzy relational model, one neuro-fuzzy modeling technique. Neuro-fuzzy modeling is a popular à ¿grey-boxà ¿ modeling technique used to model complex, non-linear plants utilizing input-output data, i.e. as an alternative to physical-based modeling. The controllable input variables of each of the generated training data set, are positioned at the centres of the fuzzy sets, so that the steady-state and dynamic performance of the model should be satisfactory whenever the control signal is stepped between the centres of its fuzzy sets. The rule confidences of the fuzzy rules are identified via the Global Least-Square (GLS) identification algorithm. The model performance is validated by using a simulated water level control system.