A predictive model for the selectivity variables of an ammonium nitrate fluidised bed granulator

Abstract

A model for the inference measurement of the selectivity variables of an ammonium nitrate fluidised bed granulator was investigated. Fluidised bed granulators are complex due to numerous processes occurring in one vessel and can thus be challenging to control. Previous studies have shown that there is no statistical significant relation between the input (operational) parameters and the selectivity (quality) variables of the fluidised bed granulator, therefore the use of artificial neural networks to estimate the selectivity variables was investigated. The objectives of the study included confirming that the selectivity variables are measured adequately and that the results of the measurements can be trusted as well as the evaluation of artificial neural network configurations as predictive models for the selectivity variables. The selectivity variables investigated are porosity and sphericity of the ammonium nitrate product granules. The porosity measurement used by the industry partner, namely the oil absorption technique, was compared to a newly developed technique called the optical porosity measurement technique to ensure the results are credible and repeatable. The correlation between the two techniques was satisfactory, with a coefficient of determination of 0.92 and a mean absolute error of 12.7 %. The high absolute error was due to the difference in the calculations of the two techniques; the oil absorption technique calculates the porosity of the particle based on a mass of oil absorbed whereas the optical porosity measurement technique calculates porosity based on the volume of the pores. The sphericity measurement technique was tested for repeatability of the technique and it was found that the results did not vary significantly when samples were analysed multiple times. The output values for the neural networks were the two selectivity variables investigated in the study. The prediction of the variables were done using separate neural networks. The input values for the neural networks were chosen as operating parameters that were measured online. These values are listed below, as used for phase 1 / phase 2 respectively. •Fluidising air flow rate / fluidising air pressure •Fluidising air temperature / fluidised bed temperature •Liquid ammonium nitrate density / liquid ammonium nitrate concentration •Liquid ammonium nitrate temperature •Liquid ammonium nitrate flow rate / liquid ammonium nitrate pressure •Atomising air flow rate. The first part of the evaluation of artificial neural networks consisted of using historical plant data that had to undergo multiple filtering stages before the data could be used for the training of a neural network. In this phase, different neuron configurations were tested using two training algorithms, namely Levenberg-Marquardt and Bayesian regularisation, on three datasets that were each assembled in their own way. The first dataset consisted of all the usable data acquired from the historical data, the second set contained only discrete data points with unique output values, while the third dataset contained data in constant time intervals of 10 minutes after each new output value was recorded. The efficiency of the constant-time dataset for use in an artificial neural network was also assessed for different time intervals as well as noise filtering of the input data. The best performing network in this phase was trained with the noise filtered constant-time data with a 10-minute time interval, using the Bayesian regularisation training algorithm with 100 neurons in the hidden layer. Data acquisition for the second phase of this study was based on the conclusions drawn from the first phase, which include: • Taking of samples for porosity and sphericity measurement in 10-minute intervals • Filter input data to reduce sensor noise • Use a large number of neurons in the hidden layer • Use the Bayesian regularisation training algorithm. The best performing neural network for the prediction of porosity, used the Bayesian regularisation training algorithm for the training of the network and had 61 neurons in the hidden layer. The training of the neural network resulted in a correlation coefficient of 0.97 and mean-squared error of 3.4×10-⁶. Subsequent simulation of the neural network with unseen resulted in a correlation coefficient of 0.90 and mean-squared error of 1.9×10-⁵. The neural network testing for the estimation of sphericity was done as with phase 1 where all neuron configurations were tested using both training algorithms. The best performing neural network for the prediction of sphericity has 75 neurons in the hidden layer and uses the Bayesian regularisation training algorithm for the training of the neural network. The training of the best neural network for the prediction of sphericity resulted in a correlation coefficient of 0.98 and mean-squared error of 1.6×10 ̄ ⁷. The simulation of the neural network with unseen resulted in a correlation coefficient of 0.96 and mean-squared error of 4.5×10 ̄ ⁷. For both the porosity and sphericity data, the hyperbolic tangent activation function was found to outperform the sigmoidal and linear activation functions in any choice of neural network.