Fig. 1. Picture of the kiln and electronics. around 750°C having as a superior limit of operation 1000°C. A schematic view of the kiln can be seen in figure 2. Fig. 4. Feedforward Neural Network structure. A Feedforward Neural Network (FNN) is a layered structure, which can include non-linearity. The basic element of a FNN is the neuron that is shown schematically in figure 3. where usual functions for F are sigmoidal, linear and hard limit. diagram of this type of training can be seen in figure This structure is supposed to overcome the problems mentioned for the previous type of training because the network is trained in a situation similar to the one that the NN will assume in a control situation. Direct inverse control. Direct inverse control is the simplest solution for control that consists of connecting in series the inverse model and the plant as can be seen in figure 8. Fig. 10. Output of forward model and training set (upper half) and Error (lower half). After the identification task, input and output sets of data are produced (which later will be split in to training and test sets). The NN training is accomplished with scaled sets. As can be seen the forward model follows very closely both the train and test sets. Fig. 14. Direct inverse control using as reference the training signal. Fig. 12. Output of inverse model and test set (upper half) and Error (lower half). Fig. 15. Direct inverse control using a new reference. Simulation of direct inverse control using both train and test sets shows that the model have captured enough information about the system exception made to the initial heating phase where data was only available starting from 300°C. When the accuracy of the models was considered sufficient (this was also evaluated from closed loop simulations as depicted in figure 13) the model was tested on the real kiln using the Data Logger connected to MATLAB to monitor the temperature and interface the control signal to be applied.
