Application of significance analysis to finding root causes of product defects in continuous casting of steel

Application of significance analysis to finding root causes of product defects in continuous casting of steel

Agnieszka Rodziewicz, Marcin Perzyk

Institute of Manufacturing Technologies, Warsaw University of Technology – Narbutta 85, 02-524 Warsaw, Poland.



The aim of this paper is to present an application of the input variable significance analysis to finding probable causes of product defects occurring in continuous casting (CC) of steel. The research was carried out using production data routinely recorded in one of Polish steel plants and basically referred to defective fraction of billets per heat as the process output. The data did not include the cases with zero defects which made the analysis difficult. The process inputs included eight parameters of different nature (physical, organizational and human). For determining which of the process input parameters are crucial for the output and which of them can be easily eliminated in further analyses two different approaches were applied and compared. The basic tool was an MLP-type Artificial Neural Network in which the relative significance was defined as the sum of the absolute weights of the connections from the given input node to all the nodes in the first hidden layer.  As a complementary method the one-way analysis of variance (ANOVA) was utilized in which the value of the F-statistics is used as a measure of the input significance. It was found that the both methods indicate that the start-time of the CC process is the factor highly influencing the fraction of defective products. The process physical parameters which are expected to have a large influence on the billet quality, i.e. deviations from nominal casting temperature and deviation from nominal casting speed also appeared to be significant, moreover their variations also highly depend on the start-time of the CC process. The final conclusion is that the direct cause of the defective products are incorrect adjustments of the casting speed occurring mainly in the morning hours, however not correlated with particular operators. This finding can considerably facilitate the identification of the root cause of the defects by the plant engineers. Some recommendations concerning the future work are also given.

Cite as:

Rodziewicz, A., Perzyk, M. (2016). Application of significance analysis to finding root causes of product defects in continuous casting of steel. Computer Methods in Materials Science, 16(4), 187 – 195.

Article (PDF):


Significance analysis, Artificial neural networks, Analysis of variance, Continuous steel casting process


Camisani-Calzolari, F.R., Craig, I.K. & Pistorius, P.C., 2003,Quality prediction in continuous casting of stainlesssteel slabs, Journal of The South African Institute ofMining and Metallurgy, 103, 651-665.

Masters, T., 1993, Practical neural network recipes in C++,Academic Press, San Diego.

Perzyk, M., Biernacki, R., Kozłowski, J., 2008, Data mining inmanufacturing: Significance analysis of process parameters,Journal of Engineering Manufacture, Proceedingsof the Institution of Mechanical Engineers, part B, 222,1503-1516.

Perzyk, M., Kochański, A., Kozłowski, J., Soroczyński, A.,Biernacki, R., 2014, Comparison of data mining toolsfor significance analysis of process parameters in applicationsto process fault diagnosis, Information Sciences,259, 380-392.

Perzyk, M., Rodziewicz, A., 2016, Application of time-seriesanalysis for predicting defects in continuous steel castingprocess, Archives of Foundry Engineering, 16, 4,125-130.

Wieczorek, T., Golak, S., 2004, An algorithm of knowledgeextraction from trained neural networks, Advances insoft computing, eds, Klopotek, M.A, Wierzchon, S.T.,Trojanowski, K. Springer Verlag, Berlin Heidelberg,New York, 470-475.