Abstract: A novel soft sensing method was proposed based on independent component analysis (ICA) fortified with variables grouping and ensemble learning method.First,the import process variables were grouped using ICA algorithm,developing multiple variable-group subspaces.Then,the coupling relation among variables and variable groups was reduced by re-sampling the data samples in the variable-subspaces.And prediction sub-models were constructed based on kernel partial least square method (KPLS).Finally,Bayesian inference method was used to integrate the outputs from sub-models,and to get the predicting results.The superiority of the proposed method was demonstrated with comparative studies of multiple soft sensors using the industrial rubber mixing process data.