The　problem　of　misalignment　of　the　original　measurement　model　is　caused　by　the　difficulty　in　obtaining　the　labeled　sample　and　the　change　of　working　condition　during　the　operation　of　the　wet-type　ball　mill.　In　this　paper,　we　introduce　a　domain　adaptive　random　weight　neural　network　(DARWNN),　thus　a　small　number　of　labeled　samples　in　the　working　condition　combined　with　the　original　working　condition　samples　can　be　used　to　implement　transfer　learning.　The　DARWNN　network　can　solve　the　problem　of　machine　learning　in　different　working　conditions,　however　it　considers　only　the　empirical　risk　but　not　the　structural　risk.　Thus　the　generalization　performance　is　poor　and　the　prediction　accuracy　is　low.　On　this　basis,　we　propose　a　domain　adaptive　manifold　regularization　random　weight　neural　network　(DAMRRWNN)　in　terms　of　manifold　regularization　to　maintain　data　geometry　structure,　so　as　to　improve　the　performance　of　the　corresponding　model.　Experimental　results　indicate　that　the　performance　of　the　proposed　methods　is　superior　to　or　at　least　comparable　with　the　existing　benchmarking　methods　and　that　the　proposed　methods　can　effectively　improve　the　learning　accuracy　of　DARWNN　and　solve　the　problem　of　soft　sensor　for　wet　ball　mill　load　parameters　under　multiple　loading　conditions.　Copyright　©　2019　Acta　Automatica　Sinica.　All　rights　reserved.