Reduction of secondary variables on soft-sensing model based on similarity in feature subspace

A novel method based on partial least-squares regression (PLS) and false nearest neighbor method (FNN) is proposed on select the most suitable secondary process variables used as soft sensing inputs. In the proposed approach, the PLS can be employed to overcome difficulties encountered with the existing multicollinearity between the factors. In the new orthogonal space, it is inspired by FNN that interpretation of primary variable would be estimated by calculating the variables´ relativity in the low-dimensional space to select secondary variables. The least square method can be used to get a soft-sensing model after the variables selection. The variable selection results of structure sample demonstrate that the method is effective and suitable for variable selection.