Identificação neuro-fuzzy aplicada a um sistema de nível de líquido

Abstract

It is not always easy to model systems mathematically, making their analysis difficult and motivating the development of different modelling techniques. In this context the area of system identification appears, providing techniques to identify systems which do not involve straightforward modelling. Identification algorithms use input and output data from the plant to be modelled to adjust the parameters of the resulting model. This work aims to identify a tank system simulated based on a real plant by using the ANFIS algorithm, which possesses characteristics both from artificial neural networks and from fuzzy systems. Its network structure allows the application of the backpropagation algorithm for training, generating a fuzzy system. The ANFIS model is non-linear, being adequate to model systems for which linear models are not appropriate. Firstly, the plant was modelled mathematically. After modelling it, ANFIS was used to identify the system. In order to do this, a random signal was generated and subsequently applied to the simulated plant, generating input-output data. By collecting these data, they were used by the ANFIS algorithm to generate a fuzzy system which had the desired characteristics. New input signals were applied to the simulated plant and to the ANFIS model, and the results showed that the identified model was able to capture both the system's dynamic and its stable state. In the end, an input signal was applied to the ANFIS model and to the mathematical model, showing that their dynamic behaviour was close, but that their stable states had different values, indicating the superiority of the identified ANFIS model.