A hybrid input variable selection method for building soft sensor from correlated process variables

Selection of the most relevant input variables for an inferential predictor is important for good prediction ability. A hybrid variable selection method is proposed for selecting input variables for support vector regression (SVR) model. The proposed method combines Taguchi's experimental design method with backward elimination method to select the most relevant variables from a large set of process variables. Taguchi's design of experiment (DoE) method was used to screen variables, as process variables are highly correlated this poses difficulty to fill in the design matrix of Taguchi's DoE method. The proposed method makes several modifications to Taguchi's method to deal with this problem. Subsequently backward elimination method was used to select the final set of input variables. The efficacy of the proposed methodology is demonstrated on an industrial case study.