As　a　key　role　in　the　future　network　functionalities,　artificial　intelligence　(AI)　has　been　studied　widely.　To　match　the　envisioned　traffic　evolution　trend　and　communication　requirement,　Network　Data　Analytic　Function(NWDAF)　is　a　new　proposed　5G　component　to　provide　analytics　for　any　Network　Functions　(NFs).　Considering　sending　all　data　to　a　central　NWDAF　instance　is　extremely　time-consuming　and　it　raises　concerns　about　security　vulnerabilities　and　data　overload,　it＇s　expected　to　design　a　distributed　NWDAF　architecture　which　is　able　to　enhance　NF　data　localization,　improve　security,　reduce　control　overhead　during　model　training,　shorten　the　training　time,　and　finally,　enhance　the　accuracy　of　the　trained　models　by　virtue　of　local　testing　on　a　real-time　network.　Federated　learning　is　considered　in　future　NWDAF　design,　while　its　detailed　structure　and　processes　are　not　standardized　yet.　This　paper　studies　the　effectiveness　of　distributed　federated　learning　based　traffic　characteristics　analysis　for　NWDAF.　Simulation　shows　that　the　learning　accuracy　is　nolinear　with　data　loss　probability.　Under　the　simulation　assumption,　the　solution　could　endure　30%　data　loss,　while　the　learning　accuracy　and　the　retransmission　dramatically　rise　when　data　loss　probability　is　greater　than　45%.