In　wireless　sensor　networks　(WSNs),　nodes　fault　diagnosis　is　the　important　measure　for　the　continuous　monitoring　service.　Efficient　and　accurate　methods　for　diagnose　nodes　fault　is　a　hot　research　topic　in　the　field　of　WSNs　in　the　recent　years.　This　paper　propose　an　improved　distributed　fault　diagnosis　method　namely　Cross-Space　Distributed　Fault　Diagnosis　(CS-DFD)　which　follows　the　DFD　algorithm,　building　high-dimensional　vector　space　model　through　the　collected　information　from　each　node,　establishing　the　cross　type　of　sliding　window　through　its　historical　data　and　the　data　from　neighbor　nodes,　setting　the　cross　direction　customizable　weights　of　fault.　Ultimately,　achieving　the　goal　of　fault　diagnosis　by　detecting　abnormal　vector　and　threshold.　The　experimental　results　show　that　it　has　reduced　the　amount　of　calculation;　simplified　harsh　conditions　of　fault　judgments;　reduced　the　consumption　of　electricity,　etc.　which　in　DFD　algorithm.　Comparing　with　the　traditional　DFD　algorithm,　the　fault　diagnosis　accuracy　improved　by　5.14%,　the　fault　of　false　alarm　ratio　(FAR)　decreased　by　2.01%　and　the　time　of　diagnosis　has　shortened　obviously.