Classification of blast furnace probe temperatures using neural networks

Abstract

Classification of spatially distributed measurements in industrial processes is often a difficult task. A typical example is given by the interpretation of the gas temperatures from horizontal probes in blast furnaces. This paper presents a neural network‐based classification of temperature measurements from an above‐burden probe. An expert on blast furnace supervision and control first classified a large set of temperature profiles into six different stereotype patterns. Roughly 40 % of this material was used for training feedforward neural networks, while the remaining profiles were used for evaluation of the networks' performance in order to determine an appropriate network size. In general, consensus was found among the networks in that similar erroneous or inconsistent classifications made by the human expert were detected. However, the network size clearly affected the quality of the classifications. The work demonstrated the merits of a rapid, consistent, and automatic (neural) interpretation, as well as the risk of misclassification and the subjectivity involved when human experts have to evaluate complex patterns.