Multiphase　characteristics　and　auto-and-cross　correlation　between　process　variables　are　two　critical　issues　to　be　addressed　for　fermentation　process　monitoring.　A　method　namely　multi-phases　dynamic　PCA　is　proposed　to　investigate　these　two　non-negligible　issues.　We　introduced　Affinity　Propagation　(AP)　clustering　to　solve　the　problem　of　unequal　length　in　batch　process,　and　to　optimize　the　clustering　number　of　AP　method,　Silhouette　index　is　applied.　After　phase　division,　autoregressive-PCA　(AR-PCA)　model　is　structured　to　remove　auto　and　cross　correlation　in　each　phase.　Model　order　is　determined　using　Akanke　information　criterion　(AIC)　and　the　model　parameters　are　estimated　by　PLS　method.　Additionally,　to　improve　the　fault　detection　effectiveness,　normal　historical　data　is　introduced　to　structure　AR-PCA　model　when　monitoring　the　new　batch.　The　experiment　result　of　penicillin　cultivation　process　indicates　the　feasibility　and　effectiveness　of　the　proposed　method.