Soft　sensors　have　been　widely　used　in　industrial　process　monitoring.　The　core　of　monitoring　complicated　industrial　processes　is　to　recognize　multi-modes　and　strategically　apply　different　sub-model.　This　paper　proposes　a　new　soft　sensor-modeling　method　based　on　BPC　(Border-peeling　Clustering)　and　PLSR　(Partial　Least　Square　Regression).　Moreover,　BPC　is　robust　towards　noise　and　sample　unbalanced　problems.　By　iteratively　peeling　off　layers　of　points,　the　cores　of　the　latent　clusters　are　revealed,　which　indicates　the　different　operation　modes.　The　three-phase　flow　process　case　proves　the　effectiveness　and　superiority　of　the　proposed　method.　Experimental　results　of　three-phase　flow　show　that　the　mean　square　error　of　the　proposed　method　is　34.1%,　which　is　better　than　other　methods.　©　2021　IEEE.