Comparison of selection methods of explanatory variables in PLS regression with application to manufacturing process data

A large number of variables are used to describe manufacturing processes in the oil, chemical and food industries. In order to pilot and optimise these processes, the manufacturer or the researcher needs both very explanatory and good predictive models of explained variables (the responses), based on reduced numbers of pertinent explanatory variables. To achieve this goal, it is therefore necessary to have access to efficient selection methods of explanatory variables. Several variable selection methods have been compared in the context of PLS regression, under the same conditions, on several real datasets of chemical manufacturing processes. Their effectiveness, evaluated on the basis of several criteria, are compared with the final PLS model for each dataset. In conclusion, we propose a stepwise variable selection based on the maximum Qcum2 criterion, similar to the Stone–Geisser index, depending on the number of eliminated variables.