Estimating Material Properties for Process Optimization

A neural network approach to the problem of estimating physical properties of a material based on the material’s chemical composition is presented. The network consists of sigmoidal hidden units and a linear output unit arranged in a feedforward architecture. As a component of a process optimization system which is applied in production processes with a priori unknown and eventually drifting characteristics, robust and fast on-line adaptation of the network is required. Therefore, a permanently updated, stack-like organized training data set and a line-search procedure for adjusting the network weights is employed. A first application has been the estimation of the “relative yield stress” of different steel qualities, which is necessary for optimizing the rolling process at a hot line rolling mill. Compared to the current state-of-the-art method a reduction of the average estimation error of about 35% has been achieved.