RBF neural network improvement for plant data based learning

Neural networks, with sufficient number of hidden layers and their nodes, can approximate any continuous function with arbitrarily desired accuracy. Their learning algorithms are constantly improving, therefore neural networks are being increasingly used for nonlinear process modeling. To obtain good process model, it is extremely important that training data is informative enough. Problem with neural networks is their sensitiveness when it comes to correlated and noisy training data. Since this is always the case with actual plant data, plant data based process model building is a challenging task. In this paper, RBF neural network upgrade with PLS method for projection into a latent subspace is described. Due to this upgrade, plant data are not used directly for neural network training, but is previously transformed using PLS method. This way, uncorrelated and less noisy training data is provided. RBF neural networks with Gaussian activation functions are used. Both RBF neural network upgraded with PLS method and standard RBF neural network are trained on actual process data. Learning improvements of upgraded RBF network are analyzed.