Adaptively changed winning number LVQ for constructing an accurate control model from enormous and low quality plant data

Construction or tuning of control models is often done using data obtained from an actually working plant. We discuss how to improve these plant data from the viewpoints of decreasing their size to a manageable number without losing their statistical property, excluding ill-suited ones, and dissolving the partial distribution to obtain an accurate control model. We call our procedure plant data purification. First, the learning vector quantization (LVQ) is improved to obtain the desired number of purified data, where, under quantization, the number of winning quantization vectors is changed adaptively and abnormal data determined by similarity with their nearest quantization vector are excluded out of a set of quantized plant data. Then the developed method is further improved to dissolve the partial distribution of data to obtain a uniform distribution. Finally, the proposed method is applied to the construction of a control model used in a continuous galvanizing plant and its effectiveness is demonstrated.