<Continue> 

2.    Accurate parameter prediction

In order to achieve optimal plate shape, two types of logics should be well developed: (1) those for accurate parameter prediction as basis for the scheduling; and (2) logics to create draft schedule based on the predicted parameters. In the parameter prediction, force calculation is the primary issue because the temperature is usually determined based on the predicted force. Due to the difficulty to measure the temperature to use it for learning, it is usually more accurate to adapt rolling force in achieving accurate force prediction, and thus to calculate temperature based on predicted force.

Prediction Accuracy Issues

During rolling there is dynamic recrystallization, and in the inter-pass period, there are static and metadynamic recrystallization. Traditionally, steel rolling was carried out in a high temperature, so after each pass, the steel could fully recrystallize and thus the strain from the previous pass was eliminated. However, in today?s rolling practice, especially with the controlled rolling and use of microalloys, a great number of passes are rolled below the recrystallization temperature (technically it is cold rolling). Due to the incomplete recrystallization, a significant portion, sometimes 80% [1], of the pass strain can be retained to the next pass. This can cause considerable error in the strain and thus the flow stress for the later passes. Currently almost every Level 2 model does not have sufficient handling to the retained strain. The error caused in this way cannot be removed by the adaptive learning. On the other hand, the grain refining through controlled rolling or other physical-metallurgical processes (e.g. precipitation) also changes flow stress which needs complicated metallurgical model to calculate.

There are also modeling, especially learning logical issues. Roll force learning is through flow stress. After each qualified piece is rolled, flow stress learning factors C₁, C₂, C₃ and C₄, denoting the coefficients of material, temperature, strain and strain rate, respectively, are recalculated. It is based on various learning fits: FIT2 (with C₁ and C₂), FIT3A (without C₄) and FIT3B (without C₃) and FIT4 (with all).

Blind adaptive learning has certain limitations. In this case, the coefficients C₃ and C₄ are actually interactively affected. For the same model grade, for example, some pass has C₃=1.2 and C₄=0.02 while another pass has C₃=0.02 and C₄=1.0, and both passes may have good force prediction. However, combining this two learning results would produce a poor model. Therefore, this kind of blind learning can only lead to limited model accuracy. Modeling guidelines should be a great help for better learning and thus higher force prediction accuracy.

Many Level 2 models fail to provide guideline for the learning factors. Some even have logical errors, for example, among all the learning fits, if the coefficient C₃ is not used for learning, the system set C₃ to zero. This means that the contribution of the strain to the force is pushed to the strain rate factor C₄ (and other parameters). This causes C₄ to fluctuate and thus hurt the long-term learning. Solution is proposed to replace the zero with a medium value (C_3m or C_4m), if a coefficient is not used for learning. This leads to the Guided Two-Parameter Learning.

The Guided Two-Parameter Learning

The Guided Two-Parameter Learning, by using carefully designed medium values for C₃ and C₄, and performing adaptation by adjusting C₁ and C₂, would be a simple but very effective solution for all kinds of issues mentioned above, and particularly for improving existing Level 2 models. This fitting mechanism is referred to FIT2G. Primary features are described below.

1. Resolves the issue that the effect of strain and/or strain rate is ignored in force prediction. Carefully designed C₃ and C₄ precisely consider the effects of the strain and strain rate.
2. Removes the limitation of adaptive learning, e.g. due to dependence of C₃ and C₄ on each other. The model can thus continue to improve itself.
3. Considers metallurgical effects. The complicated metallurgical effects can be fully represented by a large number of well-designed flow stress coefficients, often 6000-18000 sets. This is an elegant and effective way to integrate metallurgical principle into Level 2 system.
4. Requires very few modifications to the source code, which is significant for improving exist Level 2. It eliminates the concern of introducing new errors into the system.

This procedure requires very little online handling to the existing production system, leaving large amount of work to the offline design. Therefore, the project risk is low. In the case of designing new Level 2 model, the solutions to the learning logics and limitation of adaptive learning, etc., can be integrated into the learning logics design, so the result will be good as well. Metallurgical effects are covered mainly based on offline design; this is good for simplicity and ability to continued upgrade in metallurgical model without the need for software change. In either case, the model is accurate, with high calculation speed, and with narrow temperature range for learning.

Application of FIT2G for Force Prediction

Application of the FIT2G may follow the steps listed below:

• Level 2 grade list creation
• Level 2 model grade list creation
• Temperature region determination
• Design of flow stress coefficients for all the model grades in all temperature region
• FIT2G learning by using designed data (e.g. source code modification)

Examples of the application can be seen in the section "PLATE MILL APPLICATION EXAMPLES" in this paper, or the recent AISTech paper [2].

<To Be Continued>

<Previous> | <Next>

|  Pt1  |  Pt2  |  Pt3  |  Pt4  |  Pt5  |  Pt6  |  Pt7  |  Pt8  |  Pt9  |  Pt10  |