Mineral Wool Production Monitoring Using Neural Networks

Homogeneity of the primary layer in mineral wool production process is required for high quality products. State-of-the-art measurement techniques for the evaluation of primary layer homogeneity are very slow and can only be applied after the product is manufactured. We present here a method that enables on-line monitoring and control and is based on experimental modeling using neural networks. The experimental method is based on image acquisition and image processing of the mineral wool primary layer structure. As a estimator of the mineral wool primary layer structure and quality, the weight of the primary wool layer is used, measured by an on line weighting device in four locations of the conveyor belt. The instrumentation of on line weighting device was upgraded for the purpose of the present experiment and enabled high speed acquisition of all measurement channels. The structure of the mineral wool primary layer was measured by visualization of the modified entrance to the on line balance using a CCD camera. All data channels were simultaneously sampled. Radial basis neural networks are used for prediction. The structure of the mineral wool primary layer is predicted on the basis of experimentally provided weights data. The learning set consists of weights images pairs. The prediction of the mineral wool primary layer structure consists of providing only weights. A good agreement between statistical properties of measured and modeled structures of the primary wool layer like spatial homogeneity of the primary mineral wool layer thickness, is shown. The results of the study confirm that the time delayed vector of weights bears enough information for the monitoring of the production process. The modeling of primary mineral wool structure is of lesser quality due to high dimensionality of the modeled variable. Keyword: mineral wool, radial basis neural networks, prediction, visualization.