In　industrial　process,　data-driven　soft　measuring　model　based　on　easy-to-measure　process　variables　can　be　used　to　inference　and　estimate　difficulty-to-measure　quality　or　quantity　parameter　effectively.　Normally,　there　is　strong　collinearity　among　these　input　features.　Moreover,　only　small-size　useful　input/output　data　pairs　for　modeling　such　difficulty-to-measure　predicted　parameters　can　be　obtained.　In　this　paper,　a　new　selective　ensemble　(SEN)　modeling　approach　based　on　variable　importance　of　projection　(VIP)　index　is　proposed　to　address　such　data.　The　VIP　values　of　different　input　features　combined　with　prior　knowledge　is　used　to　make　feature　selection.　These　selected　features　are　used　to　construct　soft　measuring　model　based　on　＇Resample　training　sample＇　ensemble　construction　strategy　and　SEN　kernel　latent　structure　algorithm.　Simulation　results　based　on　mechanical　frequency　and　Near-infrared　(NIR)　spectral　data　show　effectiveness　of　the　proposed　method.　©　2018　IEEE.