Improved　local　tangent　space　alignment　(ILTSA)　method　with　adaptive　neighborhood　selection　was　presented,　aiming　at　solving　the　problem　of　over-high　dimensions　and　redundancy　in　the　mixed　fault　feature　set.　As　traditional　neighborhood　selection　method　was　not　applicable　to　the　varied　curvature　and　non-uniformly　sampled　manifold,　to　keep　the　local　linearity　by　considering　the　sample　density,　the　local　curvature　and　the　deflection　angle　of　local　tangent　space,　method　of　selecting　the　neighborhood　adaptively　was　proposed　to　improve　the　robustness　of　the　algorithm.　An　improved　Fisher　criterion　method　of　feature　selection　was　also　proposed　to　improve　the　accuracy　of　fault　diagnosis.　Firstly　the　low　redundant　features　were　selected　to　make　the　high　dispersion　between　classes　and　low　dispersion　within　a　class.　Then　the　sensitive　features　were　compressed　to　reduce　dimensions　with　the　ILTSA　method.　Finally,　the　feature　subset　was　fed　into　the　k　nearest　neighbor　classification　(KNNC)　to　identify　the　fault.　The　test　on　different　fault　position　and　severities　of　rolling　bearing　verified　the　validity　of　the　proposed　method.　©　2017,　Editorial　Board　of　Journal　of　Huazhong　University　of　Science　and　Technology.　All　right　reserved.