Modified Training and Optimization Method of Radial Basis Function Neural Network for Metrics Performance Guarantee in the Auto Association of Sensor Validation Tool

This work presents the use of radial basis function artificial neural network to estimate the sensors measurements, exploring the analytical redundancy existent among different sensors in a process. However, in order to guarantee good performance of the network the training and optimization process was modified. In the conventional training algorithm, although the stop criteria, such as summed squared error, is reached, one or more of the individual performance metrics of the neural network may not be satisfactory. The performance metrics considered are Accuracy (training error), Sensitivity matrix (sensors propagated error to the estimations) and Filtering matrix (sensor propagated noise to the estimations). The paper describes the proposed method including all the mathematical foundation. A dataset of a petroleum refinery is used to train a RBF (Radial Basis Function) network using the conventional and the modified method and the performance of both will be evaluated. Furthermore, AAKR (Auto-Associative Kernel Regression) model is used to the same dataset. Finally, a comparison study of the developed models will be done for each of the performance metrics, as well as for the overall effectiveness in order to demonstrate the superiority of the proposed approach.