PCA,　as　the　representative　of　the　traditional　global　fault　monitoring　algorithm,　is　a　linear　feature　extraction　algorithm.　So　global　algorithm　may　cause　a　lot　of　false　alarms　in　fault　monitoring　of　fermentation　process　with　nonlinear　and　multi-stage　characteristics.　In　this　paper,　the　Just　In　Time　Learning　(JITL)　strategy　is　introduced　into　the　Kernel　Principal　Component　Analysis　(KPCA)　algorithm.　The　local　model　is　established　with　the　data　similar　to　the　current　online　sample.　Because　the　local　model　can　represent　the　current　state　of　the　system,　it　is　not　necessary　to　identify　the　stages　before　monitoring.　At　the　same　time,　the　local　KPCA　model　can　be　used　to　feature　extract　nonlinear　data.　The　data　generated　by　the　PenSimv2.0　simulation　platform　is　used　for　verifying　the　algorithm.　The　results　show　that　this　method　has　a　better　effect　than　KPCA　algorithms.　©　2016　IEEE.