This research summary discusses the application of neural networks, evolutionary strategies, and other artificial intelligence techniques in the modeling and optimization of the models used for temperature, rolling force, and torque calculation in heavy plate milling.

INTRODUCTION

Heavy plate production is a highly competitive market. It requires mathematical models to determine, with the highest possible accuracy, the factors characterizing the rolling process,'-3 particularly temperature, rolling force, torque, and mechanical properties. Artificial intelligence can help in the modeling and optimization of these processes.
The strategy adopted in the plate mill rolling models is based on maximizing the reduction applied each pass without exceeding the physical limits of the mill or hindering the desired metallurgic properties of the product. This philosophy is aimed at minimizing the overall rolling process time, thus maximizing the output while ensuring the quality of the product.
Most models are based on mathematical formulae, which use empirical constants that must be adapted to the characteristics of the mill in which the system is implemented. These constants must be readjusted periodically to adapt them to any structural, mechanical, or operational changes. This calibration entails the analysis of huge volumes of data from the historic distortions and errors in the signals received from the various sensors.

THE INDUSTRIAL ENVIRONMENT
Heavy plate production is ahot rolling process. Mills usually are equipped with a single four-high reversing stand, used to produce heavy plate from continuously cast slabs. Slabs are charged into the reheating furnaces. After the required residence time, the slab to be rolled is discharged from the furnace onto the roller tables of the mill. Slabs up to 230 mm thick are normally discharged from the furnace at a temperature of 1,215-1,230 deg C, and for 280 mm thick slabs, the temperature is approximately 1,160 deg C. When a slab exits the furnace, a pass calculation algorithm predicts the sequence of rolling operations required and the characteristics of the plate after each pass. This calculation works backward-starting from the final characteristics of the rolled plate (dimensions and final temperature), the model calculates the rolling passes, going back to the first pass. First, it is necessary to predict the final rolling temperature. Then, the model calculates the temperature drop in each pass and the rolling force and torque conditions required. In the rolling phase, the plate temperature, force, and torque required to produce a given reduction profile are calculated for each pass. The rolling stand makes the number of passes required to obtain the plate dimensions requested by the customer.
Three different types of rolling practices are considered:
* As-rolled. In this process, the target is the dimension of the plate. The plate temperature after the last pass is around 950 deg C.
* Normalized rolling. In this process, the plate is heat treated after the rolling, repeating the plate in a furnace to obtain the desired mechanical properties.
* Controlled rolling. In this process, the target is not only the dimension of the plate but also its finishing temperature, aimed at obtaining a smaller grain size. There are two modes: soft annealed rolling (final temperature around 885-895 deg C) and hard annealed rolling (final temperature of 730-750 deg C). The properties of plate depend on both chemical composition and rolling practice.
Steel is classified based on chemical components: carbon steels (classified by their proportion of carbon content), alloy steels (containing significant amounts of alloying elements), and high-strength low-alloy steels (known as HSLA, with a low carbon content and a microstructure consisting of fine-grain ferrite as one phase and a hard second phase of martensite and austenite). Alloy steels, which are usually made more carefully than carbon steels, are used where strength, hardness, creep and fatigue resistance, and toughness is required.

CONCLUSION
Optimizing the empirical constants for the mathematical models does not suppose big changes in the actual models. Unfortunately, the optimizing methods based on ES are limited by the accuracy of the mathematical models. If the model has a correct physical correspondence, then the results are accurate. Likewise, if the physical model is wrong, the results will not be good enough. In that case, the ANN black-box model, which can provide excellent results, is recommended. The ANN methods, which are independent of the existing mathematical models, can consider other variables not included in the mathematical models. However, replacing the actual models with ANN black-box models is not free of problems. The training process for ANN is difficult, and people are distrustful when blackbox models are used because they cannot explain the output.
For models with a solid physical base, the best results were given by tuning their parameters with the ES methods. For final temperature adjustment, replacing the model was a better solution. In force and torque models, the adaptive models are absolutely necessary but implemented as ANN.