Truss　structure　is　widely　used　in　civil　engineering　applications　for　its　advantages　of　easy　transportation,　convenient　assembly　and　uniform　loading.　However,　it　is　difficult　to　achieve　realtime　health　monitoring　because　of　connection　diversity　and　complexity　of　truss　structures.　As　a　novel　structural　health　monitoring　technique,　electro-mechanical　impedance　method　could　monitor　the　health　state　of　one　structure　by　measuring　the　spectra　of　impedance　or　admittance　of　the　piezoelectric　elements,　which　are　bonded　on　the　surface　of　this　structure.　This　approach　has　the　advantages　of　nonparametric　model　analysis,　easy　sensor　installation　and　high　local　sensitivity,　especially　in　sensitive　frequency　range.　The　damage　information,　which　is　tested　and　recorded　by　using　electromechanical　impedance　method,　could　convert　into　intuitive　results　through　neural　network　because　of　its　good　ability　for　nonlinear　mapping.　In　this　paper,　a　three-layer　assembly　truss　structure　was　chosen　as　experimental　object,　piezoelectric　elements　were　bonded　on　structure　joints　to　measure　structural　impedance　spectra,　the　change　of　these　structural　impedance　spectra　was　tested　and　recorded　under　high　frequency　excitations　when　different　truss　bars　were　loosed,　and　then,　one　back-propagation　(BP)　neural　network　was　built　and　trained　by　this　damage　information,　which　were　treated　as　input　samples.　These　results　show　that　the　sensitivity　of　impedance　method　is　not　the　same　to　different　frequency　range　and　trained　neural　network　could　quickly　identify　loosen　truss　bars.　©　(2013)　Trans　Tech　Publications,　Switzerland.