The　frequency　response　functions　(FRFs)　are　frequently　used　to　identify　dynamic　properties　of　the　mechanical　joints.　However,　the　ill-posed　problems　of　the　system　are　caused　by　the　contamination　of　measurement　noise　even　if　it　is　processed　by　reduction　technique.　In　order　to　solve　this　problem,　put　the　noised　FRFs　into　identification　equations　to　theoretical　derivation,　explained　the　degree　of　influence　on　the　identified　results　by　noised　FRFs　in　theory.　In　order　to　make　the　identified　results　represent　real　properties　of　the　joints,　an　improved　FRFs　method　was　developed　based　on　subtraction　term　frequency　selection,　in　which　the　measured　data　were　processed　could　get　high　degree　of　anti-noise　frequency　before　parameters　identification.　This　method　avoids　the　occurrence　of　ill-posed　problems.　The　results　of　simulation　show　that　the　proposed　method　can　solve　the　ill-posed　problems,　thus　significantly　improve　the　accuracy　of　identification.