DATA MINING METHODS IN HOT STEEL ROLLING FOR SCALE DEFECT PREDICTION Jarno J. Haapamäki, Satu M. Tamminen and Juha J. Röning Neurogroup / ISG University of Oulu, Department of Electrical and Information Engineering Computer Engineering Laboratory Linnanmaa PO BOX 4500 90014 UNIVERSITY OF OULU FINLAND jjhaa@ee.oulu.fi, satu@ee.oulu.fi, jjr@ee.oulu.fi ABSTRACT Scale defects are common surface defects in hot steel rolling. The reasons for such defects are not straightforward. With data mining methods, the multidimensional dependencies between process variables and product composition can be discovered. For this research, a high-dimensional data set from Rautaruukki Oyj, Raahe, Finland was gathered. The data contained both averaged values and process values measured with different frequencies. The synchronisation of the variables as well as the allocation of the measurements on the steel strip were solved before the modelling phase. The research enabled the visualisation of the rolling process and scale defect modelling. Self- organizing maps (SOM) were used for these tasks. KEY WORDS data mining, neural networks, hot steel rolling, scale defects 1. Introduction Scale defects are a common group of surface defects in hot steel rolling, and it is therefore important to recognize the risk factors that cause scale on the surface of steel products. Hot-rolled steel strips were chosen for this study, because of their stringent surface quality requirements. Oxidation of steel leads to a three-layer scale consisting of wüstite, FeO, magnetite Fe 3 O 4 and hematite Fe 2 O 3 . Wüstite, which is the innermost layer, is stable only at temperatures above 570°C. In hot rolling process conditions it constitutes roughly 95% of the scale [1,2,3]. There are several scale types with different mechanisms of formation. Rolled-in or black scale develops when harder oxides are rolled into [1] the surface during the finishing process. Red scale is mainly associated with a high Si-content, although that is not a necessary condition [2]. In red scale detection, there is only a small possibility for confusing red scale with some other defect [4]. When the two types coincide, it is possible that some rolled-in scale defects will be ignored by the detection system, and therefore strips with over 0.1% Si-content were investigated only for red scale and the other part of the data for rolled-in scale. The origin of scale defects has been a topic of interest for many research projects, but it is still hard to find literature on the modelling of defects. No physical model for scale formation has been formulated so far. The problem for modelling is due to high dimensional variable group with their interactions. Several reasons for scale formation are mentioned in the literature, including the effects of temperature and time [1,3,5,6], rolling forces and reduction [1,6], steel composition [7] and gas atmosphere. Silicon content and reheating temperature [3] are also relevant factors, since molten fayalite, Fe 2 SiO 4 , accelerates the scaling rate [2]. Furthermore, uneven F1 F6 Coilbox Crop shear pyrometer F6 pyrometer Surface inspection system scale breaker Figure 1. Finishing mill layout.