Application of Association Rule Mining to Detect “Spike” Disruptions in Aluminum Production

Abstract

The article focuses on applying association rule mining to predict the "Anode spike"-type process disruptions using daily average monitoring data from a series of reduction cells in the experimental area of the Sayanogorsk Aluminum Smelter. The data were binarized according to different criteria for grouping the values of the process parameters into ranges: statistical norms, quartiles, and ranges which are attributed to the occurrence of disruptions. Prediction models were built as a set of association rules. The quality metrics aided in defining the optimum parameters to be used in the model settings, with the results of its validation presented as well. The model selected for the implementa-tion uses binarization based on quartiles. The resulting validation values suggest that the model is effective enough for practical use.