A study on the causes of deviation in mechanical properties of thin steel sheets

Abstract

Multi-layer perceptron (MLP) neural networks were trained using measured process data from Rautaruukki Hämeenlinna and Raahe Works. The MLPs had process parameters as inputs and the difference between the average and the actual yield strength (ΔRp02) as output. The data set consisted of various steel grades, the average Rp02 values being calculated separately for each grade. By studying the response of the MLP to changes in input variables it was possible to draw conclusions of the causes of variation in measured yield strength. It was observed that the process parameters contain sufficient information to predict the variation and that the variables responsible for the variation differ from grade to grade. A sensitivity analysis on the variations of input variables showed that the response of MLP on the variation of one variable may depend greatly on the values of the other variables.