2.12 Solution to Continued Grade Increase and Missing Grade File
Three solutions have been planned to solve the problem of continuously increased steel grades. (1) Because the large number of models have already been designed, in most cases, even a grade with new chemical composition appears, mill engineer can operate inside the Level 3 database to link this grade with an existing Level 2 grade, so the models of the existing Level 2 grade are available for this new grade. This would expect a very low error if the chemical composition is similar. (2) Inside Level 2 system, seven defense lines have been established to solve the problem of missing grade or missing model grade. The Level 2 would search the best matched grade or model grade for the missed one. Level 2 may create a new grade file by copying the best matched grade file and change related parameters. After learning, the data will be specifically suitable for the new model grade. (3) It is discussed to supply a new software package for automatic design of grade models. This is an offline design program. When a new grade appears, user only needs to enter chemical composition, and select the suitable entries from the lists of Product Thicknesses, Process Types, Slab Thicknesses and Rolling Stages, together some data such as whether there is a hold, and if any, the resume temperature. The package would immediately create the model grade list and design model data, and generate a grade file for every model grade. This package is beyond the work contract scope of the current project.
2.13 Onsite Testing
Onsite testing is a big part of a
project in China. After design and development, the system is installed into the
production system. The testing indicated a very high parameter prediction
accuracy. Operators stated that for force prediction error was reduced to mere
one half of earlier level.
It is interesting
to be mention that in China project, usually, the onsite
testing takes 50-100% of the total development time. This is
different from US project such as in Oregon Steel, in which
testing only takes very little portion.
By June 2010, major issues involved in the NISCO force project testing included: (1) for a while, the primary headache was the new grade. Often, due to communication problem among various divisions, it is hard to have early knowledge one day ahead on grades to be rolled. It often caused delay for the testing because the new model for the new grade(s) were not yet designed. (2) While the force prediction accuracy is considerably high (which can be judged by comparing the predicted and measured forces), number of passes tends to increase, and the pattern of the draft distribution among passes is not optimal. Draft schedule sometimes
was not improved; it could even get worse. It was clear that the logics for the draft scheduling have something remaining to be improved (See FURTHER DEVELOPMENT FIELDS section for further discussion).
[Later note: Draft schedule logics was then improved, and the new Level 2 model
have been running in production line since then].
2.14 Force Prediction Accuracy
Table 4 compares force prediction error levels in various occasions. NISCO抯 old model (before the improvement from Metal Pass) had much higher force prediction error than EOS’ old model. Therefore, a force model improvement is of great importance for NISCO. Since the project is not yet finally finished, error level for the new model is not yet generated, so the guaranteed error level is showed. The actual accuracy level should at least reach this guaranteed one, most likely much higher. In EOS case, even the troubled grades had reached this accuracy level. Based on observation in NISCO, it is recognized that this model is very accurate; the error from the new model is commonly only one half of the old model – this is the comment from the NISCO Level 2 operators. Force prediction error is dominantly below 5%. This new model for NISCO, with its delicate design and careful modeling based on large amount of data, is considered to be one of the world抯 finest, with metallurgical principles integrated and learning logical problems fixed. The primary difficulty in this project was not to achieve high force prediction accuracy; this was easy to accomplish. Rather, the problems mainly existed in the testing environment, such as adding new grade without advanced notice, and pass scheduling logic issues. This is also evident from the section on onsite testing.
Table 4: Force Prediction Accuracy