Calibration of an on-line analyzer using neural network modeling

Abstract

The goal of the project was to predict the ash content of raw coal in real time using the Americium-137 and Cesium-241 scintillation counts from an on-line analyzer. Rather than regression methods (that are current industrial practice), neural networks were used to map the scintillation counts to percentage ash. Quick stop training was used to prevent overfitting The noise and sparseness of the data required that the training, calibration and prediction subsets are statistically similar to each other. Therefore, Kohonen networks were first used to detect the features present in the data set. Three subsets were then built such that they had representative members from each feature. Neural network models were developed for the screened coal, the unscreened coal and the combined data respectively. The results show that the performance of the combined model was comparable to the performance with two different models for the screened and unscreened data. Due to the variance in the sample data, the neural networks (screened, unscreened and combined) did not predict individual samples well. The network predictions were, however, accurate on the average. Compared to the common regression approach, neural network modeling demonstrated much better performance in ash prediction based on certain criteria.