With　the　large　structural　health　monitoring　system　successfully　developed　and　applied,　it　has　become　to　be　focus　how　to　make　full　use　of　the　redundant　and　complementary　information　and　assess　on　structural　healthy　states.　A　new　structural　damage　identification　method　is　proposed　in　this　paper.　In　this　method,　rough　set　is　employed　to　process　initial　data　and　reduce　attributes　in　advance.　Thus　a　probabilistic　neural　network　(PNN)　is　employed　to　fuse　multi-sensor　data　and　conclude　the　damage　identification　results.　To　validate　the　proposed　method,　six　damage　patterns　from　a　7-DOF　building　model　are　identified　finally,　and　a　comparison　is　made.　The　results　show　that　the　method　can　reduce　spatial　dimension　of　data,　as　well　as　have　a　good　consist　with　un-attributes　reduction.