A Modified Algorithm for Training and Optimize RBF Neural Networks Applied to Sensor Measurements Validation

This paper presents the use of a radial basis function artificial neural network to estimate sensor readings exploring the analytical redundancy via auto-association. However, in order to guarantee optimal performance of the network, the training and optimization processes have been modified. In the conventional training algorithm, even if a stop criterion, such as summed squared error, is reached, one or more of the individual performance metrics, including: i) accuracy; ii) robustness; iii) spillover and iv) filtering of the neural network, may not be satisfactory while validating sensor measurements. Essentially, the proposed modification in the training algorithm is based on seeking to ensure that one or more of the metrics are met. This paper describes the proposed algorithm including all of its mathematical foundation. Afterward, a data set of a water injection pump for an oil and gas processing unit was used to train the RBF network using the conventional and the modified algorithm, and the performance of each was evaluated. Furthermore, the AAKR model is applied to the same dataset as a quality reference parameter. Finally, a comparison analysis of the developed models is presented for each of the performance metrics, as well as for overall effectiveness, demonstrating that the main advantage of the proposed approach is to obtain the estimation results equivalent or superior to the AAKR with shorter runtime and the disadvantage of having higher complexity during the model training.