Use of neural networks in the capacitance imaging system. Technical note

Abstract

The US Department of Energy`s Morgantown Energy Technology Center (METC) has developed a capacitance imaging system (CIS) to support its fluidized-bed research programs. The CIS uses 400 electric displacement current measurements taken between combinations of pairs of 32 electrodes to obtain a measure of the fluidized-bed material density in the volume between the electrodes. The measurements are simultaneously made for three other sets of horizontally-oriented 32 electrodes with the four sets evenly spaced vertically. This report describes the development of a method of using the 400 current measurements per level as the input to a neural network to produce the 193-pixel density estimates defined for each level. A 417-neuron subnetwork using 4,047 weights is defined as the system used to determine a set of 32-pixel densities in one of the annular regions of the fluidized-bed cross section. The same subnetwork with different values of weights is used for the other five annular regions that cover the rest of the cross section. An averaging technique is used to determine the density of the small central region. The methods used to optimize the set of weights for each of the six subnetworks are described. The results of tests using calibration electric current data as inputs to the neural system showed that these density estimates have less error than three previously developed methods of converting current measurements into pixel density maps. A comparison of the density maps produced by the neural system and the alternate three methods using input fluidization data also indicates the superior performance of the neural network approach.