For　multistage,　time-variant,　nonlinear　characteristic　and　unavailable　on-line　product　qualities　of　batch　process,　a　multi-phase　Multi-way　partial　least　squares　(MP-MPLS)　method　is　proposed.　Using　ISODATA　dynamic　clustering　algorithm,　process　data　was　automatically　divided　into　several　operation　stages　according　to　relevance.　Then,　Using　recursive　DTW　algorithm　to　synchronize　these　unequal　Sub-phase,　and　sub-phase　MPLS　models　were　developed　for　every　phases　for　on-line　monitoring　and　quality　prediction.　The　proposed　method　easily　handles　the　following　problems:　(1)　static　single　model;　(2)　process　and　its　model　do　not　match;　(3)　linear　method　may　not　be　efficient　in　compressing　and　extracting　nonlinear　process　data.　The　idea　and　algorithm　are　illustrated　with　respect　to　the　typical　data　collected　from　a　benchmark　simulation　of　fed-batch　penicillin　fermentation　production.　For　comparison　purposes,　a　traditional　MPLS　model　and　a　knowledge-based　MPLS　model　was　established.　The　results　demonstrate　the　effectiveness　of　the　proposed　method.　©2010　IEEE.