A　new　time-domain　modal　identification　method　of　ARMA　(2,1)　model　based　on　improved　EMD　is　presented　in　this　paper.　It　is　able　to　partly　solve　the　problems　in　identifying　modal　parameters,　such　as　mode　mixing　in　classic　EMD,　no　input　excitation　and　no　need　determining　the　model　order　of　ARMA　model　in　advance.　Firstly,　an　improved　EMD　(IEMD)　method　is　proposed　in　this　paper.　A　band-pass　filter　method　based　on　Fast　Fourier　Transform　(FFT)　is　as　a　pre-processer　of　classic　EMD.　Then　the　real　IMF　is　determined　by　the　correlative　coefficient　between　the　separated　IMFs　and　the　primary　signal.　Finally,　the　Natural　Excitation　Technique　(NExT)　and　ARMA　(2,1)　model　are　combined　to　identify　structural　modal　parameters.　To　illustrate　the　effectiveness　of　the　proposed　method,　modal　parameters　of　a　7　storey　steel　frame　are　identified.　The　study　results　shows　that　the　approach　proposed　can　extract　modal　parameters　effectively,　and　also　have　excellent　adaptability.