Conventional　multiway　kernel　partial　least　squares　(MKPLS)　method　needs　to　calculate　all　the　measured　variables　in　every　pair　of　two　variables　when　using　the　kernel　trick,　which　causes　great　amount　of　calculation　and　memory　requirement.　Aiming　at　the　nonlinear　and　calculation　burden　problems　in　online　quality　prediction　of　batch　process,　a　new　feature　space　(FS)　based　kernel　partial　least　squares　algorithm　is　proposed　to　carry　out　the　on-line　quality　prediction　of　batch　processes.　First,　the　proposed　algorithm　expands　the　3-D　collected　data　into　2-D　ones　and　performs　the　normalization　processing.　Then,　a　feature　vector　selection　method　is　applied　to　reduce　the　data　calculation　burden　during　PLS　kernel　trick　implementation.　Last,　aiming　at　the　blindness　of　traditional　feature　vector　selection　algorithm　in　feature　vector　selection　sequence,　the　quality　data　are　taken　into　account　and　a　new　feature　vector　selection　method　is　suggested　to　solve　the　nonlinear　problem　in　online　soft　sensing　and　further　improve　the　accuracy　of　online　soft　sensing.　Finally,　the　proposed　method　was　applied　in　the　penicillin　fermentation　process　simulation　and　the　actual　process　online　monitoring,　which　verify　the　validity　of　the　proposed　method.　©,　2015,　Science　Press.　All　right　reserved.