In　patients　with　stroke　and　spinal　cord　injury,　motor　function　is　reduced　or　even　lost　because　motor　nerve　signals　cannot　be　transmitted　due　to　nerve　injury.　Functional　electrical　stimulation　(FES)　is　one　of　the　most　important　rehabilitation　techniques　for　the　treatment　of　motor　impairment　in　patients　with　stroke　and　spinal　cord　injury,　which　has　been　widely　used　in　the　recovery　and　reconstruction　of　limb　motor　function.　In　this　paper,　we　propose　a　neural　network　based　modeling　method　and　control　implementation　of　FES　system　for　upper　limb　neurorehabilitation.　A　dynamic　neural　network　model　based　on　Hammerstein　structure　is　proposed　for　modeling　the　elbow　joint　motion　under　functional　electrical　stimulation.　A　closed-loop　control　system　for　FES　is　realized　using　iterative　learning　control　(ILC)　and　achieved　an　excellent　tracking　performance.　Both　simulation　and　experiment　are　carried　out　to　demonstrate　the　results.　Considering　the　20　tests　of　the　model,　the　average　of　average　relative　error　(ARE)　and　root　mean　square　error　(RMSE)　of　the　testing　samples　are　4.11%　and　4.12　degrees,　respectively.　The　ability　of　ILC　system　to　resist　model　disturbance　is　discussed,　and　the　maximum　error　between　the　actual　elbow　joint　trajectory　and　the　desired　trajectory　for　each　motion　cycle　is　analysed.　As　the　number　of　iterations　increases,　the　actual　elbow　motion　can　track　the　desired　trajectory.　The　experiment　verifies　that　the　real-time　system　can　realize　the　desired　trajectory　tracking.　The　results　show　that　the　established　dynamic　neural　network　model　is　suitable　for　studying　the　motion　characteristics　of　elbow　joint　under　electrical　stimulation.　It　is　feasible　to　train　the　network　with　the　aid　of　genetic　algorithm,　and　the　iterative　learning　strategy　can　achieve　excellent　control　effect　in　elbow　joint　FES　system.　(C)　2019　Elsevier　B.V.　All　rights　reserved.