Abstract

Ilmari Juutilainen (2002): Predicting the Strength of Steel Plates Using Regression Analysis and Neural Networks. University of Oulu, Department of mathematical sciences, Statistics. Master’s thesis.

The work is done for plate products department of Rautaruukki’s Steel factory at Raahe. The aim of the work is to develop a model that predicts the breaking strength of steel plates using the element concentrations of slabs and some variables of the rolling process. The model used nowadays for the same purpose is found to be defective. With the new model it is hopefully possible to achieve remarkable economical benefits when the production can be planned more accurately.

Used modelling methods are regression analysis and neural networks. Linear regression, regression with logarithm transformed response and also non-linear regression was used.

Modelling with perceptrons was also tried and the results are compared with the regression models. Special attention was paid to avoid overfitting because the data were very large and there were a lot of explanatory variables, which were correlated among each other.

In the text the basic principles of the statistical methods used are discussed. The prediction of the different models is analyzed and the effect of explanatory variables is visualized. New models are compared also to the model used nowadays. The new models can predict the strength much more accurately than the old model, so the work has filled the expectations given to it. For the best models the standard deviation of the error term was 10 MPa and the absolute mean error was 7 MPa. Better predictions were not achieved with neural networks than with regression models.