Application of on-line adaptable Neural Network for the rolling force set-up of a plate mill

This paper introduces a Neural Network application to a plate mill to improve the model's prediction ability for rolling force. Since thickness accuracy is highly related to the rolling-force precision, its improvement is very important. Conventional methods with simple mathematical models and a coarse learning scheme are not sufficient to maintain a good prediction ability for the rolling force because the rolling force variable has very nonlinear and time-varying characteristics. These problems are alleviated when an on-line adaptable Neural Network is applied instead. Basically, the Neural Network is capable of compensating the nonlinear model deficiency, and its on-line training reduces the prediction errors caused by time-varying rolling conditions. The field test at Pohang No. 2 Plate Mill has showed that the proposed method has improved the prediction ability by 30%. (C) 2004 Elsevier Ltd. All rights reserved.