Feature selection based on fuzzy clustering analysis and association rule mining for soft-sensor

Soft-sensors have been widely used for estimating product quality or other key variables. To achieve high estimation performance for soft-sensor design, it is important to select appropriate input or explanatory variables. This paper presents a new feature selection method applied to Soft-sensors. The proposed method, referred to as FCA-ARM (fuzzy clustering analysis-association rule mining). The measured variables were first clustered on the basis of the correlation by fuzzy clustering analysis, and each variable cluster was further evaluated by association rules mining, which can discover the important input variables that are related to the output variable based on the Apriori algorithm. By applying this method with the influence degree analysis, the overlap information can be effectively eliminated, and the important variables can be obtained as input variables. The usefulness of the proposed FCA-ARM feature selection method is demonstrated through an application to mechanical property forecasting in industrial hot rolling process.