Neural networks for steel manufacturing

Abstract

For several years, the Industrial and Building Systems Group at Siemens has successfully used neural networks for second level process automation in basic industries. Worldwide, Siemens currently has more than 20 neural network applications running in a dozen plants, 24 hours a day. Several aspects of neural networks contribute to their usefulness in the steel industry. First, they speed the development of new applications. In the past, steelmakers had to develop and program special analytical models, a laborious and time consuming process. Neural networks are simple mathematical structures that gather knowledge by learning from examples, which a computer can do automatically. Besides being so much quicker and easier, neural models also often achieve better performance than do analytical models in practical applications. Second, neural networks can handle highly nonlinear problems, making them vastly superior to classical linear approaches. Finally, neural network are able to adapt online. Applying our solutions to real world technical processes at Siemens required that we surmount several challenges, which involved extensive engineering effort. In particular, we needed to improve the control system without discarding existing solutions.