Improving the prediction of the roll separating force in a hot steel finishing mill
Frayman, Yakov, Rolfe, Bernard F., Hodgson, Peter D. and Webb, Geoffrey I. 2003, Improving the prediction of the roll separating force in a hot steel finishing mill, in IPMM 2003 : Intelligence in a small materials world : selected papers from the Fourth International Intelligent Processing and Manufacturing of Materials Conference, Intelligent Processing and Manufacturing of Materials, Sendai, Japan, pp. 1-12.
IPMM 2003 : Intelligence in a small materials world : selected papers from the Fourth International Intelligent Processing and Manufacturing of Materials Conference
Meech, John A. Kawazoe, Yoshiyuki Kumar, Vijay Maguire, John F.
Intelligent Processing and Manufacturing of Materials
Place of publication
This paper focuses on the development of a hybrid phenomenological/inductive model to improve the current physical setup force model on a five stand industrial hot strip finishing mill. We approached the problem from two directions. In the first approach, the starting point was the output of the current setup force model. A feedforward multilayer perceptron (MLP) model was then used to estimate the true roll separating force using some other available variables as additional inputs to the model.
It was found that it is possible to significantly improve the estimation of a roll separating force from 5.3% error on average with the current setup model to 2.5% error on average with the hybrid model. The corresponding improvements for the first coils are from 7.5% with the current model to 3.8% with the hybrid model. This was achieved by inclusion, in addition to each stand's force from the current model, the contributions from setup forces from the other stands, as well as the contributions from a limited set of additional variables such as: a) aim width; b) setup thickness; c) setup temperature; and d) measured force from the previous coil.
In the second approach, we investigated the correlation between the large errors in the current model and input parameters of the model. The data set was split into two subsets, one representing the "normal" level of error between the current model and the measured force value, while the other set contained the coils with a "large" level of error. Additional set of data with changes in each coil's inputs from the previous coil's inputs was created to investigate the dependency on the previous coil.
The data sets were then analyzed using a C4.5 decision tree. The main findings were that the level of the speed vernier variable is highly correlated with the large errors in the current setup model. Specifically, a high positive speed vernier value often correlated to a large error. Secondly, it has been found that large changes to the model flow stress values between coils are correlated frequently with larger errors in the current setup force model.
Field of Research
091006 Manufacturing Processes and Technologies (excl Textiles)