A neural network model of a fluidised bed

Abstract

A neural network was used to model experimental fluidisation data - bubble size and velocity - from a laboratory size fluidised bed column. Experiments were performed for several particle sizes and temperatures. A classical feedforward neural network trained with a backpropagation algorithm was used. Data pre-processing turned out to be vitally important. Both particle sizes and temperatures were represented as fuzzy sets. This representation took care of a smooth transition from one value to another on the input side of the network thus making training and generalisation of the network easier. After training, the neural network was capable not only to predict the velocities of bubbles or slugs but also to identify the transition from a bubbling fluidised bed to a slugging bed.