Municipal　solid　waste　incineration　(MSWI)　is　a　widely　used　domestic　waste　resource　treatment　technology.　However,　this　process　products　pollution　dioxins　(DXN).　It　has　high　toxicity　and　durability.　This　is　one　of　the　main　reasons　of　＂not　in　my　backyard＂　effect　when　constructing　MSWI　plant.　As　the　long　period　and　high　cost　of　off-line　DXN　detection　method,　it　is　difficult　to　realize　the　real-time　monitoring　of　DXN　emission　concentration.　The　massive　unlabeled　samples　in　the　industrial　field　contain　the　generation　mechanism　of　DXN,　which　are　not　fully　utilized.　Aim　at　the　above　problems,　this　paper　proposes　an　unmarked　samples　and　improved　deep　belief　network　(DBN)　based　method　to　construct　DXN　soft　measurement　model.　Firstly,　a　large　number　of　unlabeled　samples　are　added　to　the　training　input　sample　set　to　improve　the　learning　ability　of　soft　measurement　model　at　pre　training　phase.　Then,　an　energy　function　is　derived　as　the　activation　function　of　restricted　Boltzmann　machine　(RUM).　Finally,　for　the　whole　DBN,　dropout　algorithm　is　used　to　improve　the　robustness　of　the　model,　and　the　adaptive　learning　rate　error　hack　propagation　algorithm　is　used　to　fine　tune　the　weight　iteratively　at　the　fine　tuning　phase.　The　validity　and　rationality　of　this　method　are　validated　by　DXN　data　set.