Mining data from a metallurgical process by a novel neural network pruning method

Many metallurgical processes are complex and due to hostile environment it is difficult to carry out reliable measurement of their internal state, but the demands on high productivity and consideration of environmental issues require that the processes still be strictly controlled. Due to the complexity and non-ideality of the processes, it is often not feasible to develop mechanistic models. An alternative is to use neural networks as black-box models, built on historical process data. The selection of relevant inputs and appropriate network structure are still problematic issues. The present work addresses these two problems in the modeling of the hot metal silicon content in the blast furnace. An algorithm is applied to find relevant inputs and their time lags, as well as a proper network size, by pruning a large network. The resulting models exhibit good prediction capabilities and the inputs and time lags detected are in good agreement with practical metallurgical knowledge.