In　order　to　accurately　divide　the　stages　and　improve　the　sensitivity　of　fault　diagnosis,　a　new　method　is　proposed　for　the　sub-stage　separation　of　batch　process.　First,　based　on　the　changes　in　the　load　matrices,　which　reveal　the　change　of　data　direction,　rough　separation　of　operation　sub-stage　is　executed.　According　to　the　criterion　of　minimum　similarity　distances,　time-slice　matrices　are　sorted　to　realize　separation　of　operation　sub-stage　more　particularly　for　batch　process.　The　two-step　separation　method　considers　evolvement　of　the　underlying　process　behavior.　Therefore,　the　method　can　accurately　divide　the　stages.　Finally,　multi-way　principal　component　analysis　(MPCA)　model　is　established　for　each　phase　and　applied　to　on-line　monitoring　for　batch　process.　The　proposed　method　is　applied　to　penicillin　fermentation　process,　the　simulation　results　show　that　the　multistage-based　MPCA　method　has　better　performance,　compared　with　the　traditional　MPCA.