Acoustic　emission　(AE)　is　often　used　for　structural　health　monitoring　(SHM)　in　the　wide　field　of　engineering　structures　and　one　of　its　most　beneficial　attributes　is　the　ability　to　localize　the　damage/crack　based　on　the　AE　events.　The　vast　majority　of　ongoing　work　on　AE　monitoring　focues　on　geometrically　simple　structures　or　a　confined　area,　but　the　AE　source　location　strategies　are　rather　complicated　for　real　engineering　structures.　In　this　paper,　an　effective　method　for　source　localization　in　realistic　structures　is　presented　based　on　the　application　of　artificial　neural　networks　(ANN),　using　finite　element　(FE)　simulation　results　of　Lamb　waves　as　the　modelling　basis.　Pencil　lead　break　experiments　and　related　FE　simulations　on　a　steel-concrete　composite　girder　are　conducted　to　evaluate　the　performance　of　the　method.　The　identification　of　different　wave　modes　is　carried　by　comparing　alternative　onset　time　detection　methods.　Numerical　results　are　found　to　be　matching　closely　with　the　experimental　results.　To　get　a　reliable　ANN　model,　the　validated　FE　model　is　used　to　create　a　comprehensive　database　with　five　different　sensor　arrangements.　It　is　found　that　the　proposed　method　is　superior　to　the　classical　Time　of　Arrival　(TOA)　method　with　the　same　input　data.　The　results　indicate　that　using　trained　neural　networks　based　on　numerical　data　is　a　viable　option　for　AE　source　location　in　the　case　of　the　I-shaped　girder,　increasing　the　likelihood　of　design　and　optimization　of　the　AE　technique　in　monitoring　realistic　structures.　©　2020　The　Author(s)