An approach to model building for accelerated cooling process using instance-based learning

Precise mathematical modelling of accelerated cooling process (ACC) is not accessible or worthwhile due to the various compositions and gauges of plates and its high system specificity. An online modelling method for ACC is developed through combining instance-based learning (IBL) with the physical dynamical process model. When a plate comes, k plates whose material composition and operating conditions are nearest to that of current plate are selected from historical instances. Then, an approach based on locally linear reconstruction is extended to be suitable for MIMO system first, and is applied to structure the parameters of current plate's dynamical model according to the selected k plates, due to that LLR could be able to preserve locally linear topology surrounding the new pattern and it is robust to the number of historical data. To guarantee the accurate of historical instance, a correction method is developed to modify the parameters of current plate when the cooling process ends. The resulting model can predict the temperature evolutions of the moving plates with various compositions and gauges during the cooling process. Experimental studies with real industrial data in one steel company show the effectiveness of the proposed modelling approach. (C) 2010 Elsevier Ltd. All rights reserved.