APPLICATION OF NEURAL NETWORKS IN THE MODELING OF HOT ROLLING PROCESSES Antonio Augusto Gorni Materials Engineer, M. Eng., Researcher at the COSIPA Steelworks agorni@iron.com.br / www.gorni.eng.br August 1998 - ABSTRACT This work describes the application of neural networks in the modeling of hot rolling processes. This relatively new technique of Artificial Intelligence was conceived more than fifty years ago, but it only became really feasible with the arrival of low cost computer processing power. The first papers about its utilization in the hot rolling field were published about six years ago. Although the first results were promising, there is still some lack of confidence about its real performance under industrial conditions, which is preventing the exclusive use of this new tool in the modeling of hot rolling processes. However, neural networks are already being used, as a standard feature, in hybrid automation models for hot strip mills, where they calculate adjusting coefficients for theoretical models. However, continuous use of these tools and its continuous development certainly will contribute to increase the general confidence in this revolutionary method and pave the way for a more intensive application in practical cases. - INTRODUCTION The advent of revolutionary steelmaking processes, new materials like high performance polymers and ceramics, and a chronic excess of capacity production made steel market very competitive. If a steelmaker wants to keep or expand its market share, it must offer products with excellent quality at an affordable price. One of the keys to achieve this goal is the automation of the steelmaking process. In fact, this is one of the major stages of evolution in a steel plant, as it improves both process and product consistency, minimizes costs and make production control easy. All these factors promote a significant increase in the process cost/benefit ratio. The automation of hot rolling processes requires the development of several mathematical models for the simulation and quantitative description of the industrial operations involved. The main feature of the neural networks - the establishment of complex relationships between data through a learning process, with no need to previously propose any model to correlate the desired variables - makes this technique very attractive in the modeling of processes where traditional mathematical modeling is difficult or impossible. Besides that, they are almost immune to noise or spurious data. The development of neural network models is relatively quick and, in most cases, simple. Several researchers performed off-line tests on the modeling of hot rolling processes using this technique, frequently getting good results. However, practical applications of this technique in the field of hot rolling are very scarce, mainly due to the lack of confidence about its performance. This distrust on neural networks arises from many factors. First of all, only recently this technique became feasible, with the increasingly wide availability of low-cost computer power. Besides that, as the mathematical