In　this　paper,　a　new　passive　robust　fault　detection　method　is　proposed.　In　virtue　of　its　simple　topological　structure　and　universal　approximation　ability,　the　RBF　neural　network　is　utilized　in　the　system　identification　for　the　fault　detection.　The　set　membership　identification　is　used　to　calculate　a　set　of　uncertain　weights　which　describes　the　model　uncertainty.　This　set　allows　obtaining　an　adaptive　threshold　of　the　residual　which　is　next　applied　to　the　robust　fault　detection.　A　model　structure　selection　scheme　which　can　delete　the　redundant　hidden　nodes　is　proposed　to　reduce　the　conservatism　of　the　uncertain　set.　A　narrower　threshold　can　be　generated　owing　to　the　contraction　of　uncertain　set　and　therefore　the　fault　detection　sensitivity　is　increased.　Three　examples　show　the　satisfying　performance　of　the　proposed　robust　fault　detection　method.　(C)　2013　Elsevier　Ltd.　All　rights　reserved.