For　the　design　of　the　fuzzy　neural　network　architecture　and　the　deficiency　of　fuzzy　sets　on　semantic　description,　a　self-organising　T-S　fuzzy　Elman　network　(SOTSFEN)　based　on　extended　Kalman　filter　(EKF)　is　proposed,　and　the　training　algorithm　is　derived.　Furthermore,　recursive　least　square　(RLS)　and　EKF　are　used　to　update　linear　and　non-linear　parameters　respectively.　Then　the　criterion　of　rule　generation　is　given　and　error　ratio　reduction　(ERR)　is　regarded　as　the　fuzzy　rule　pruning　strategy.　Finally,　the　simulation　results　of　system　identification　and　sewage　treatment　modeling　show　that　the　precision　and　generalization　ability　of　SOTSFEN　are　ensured,　and　a　simpler　architecture　network　can　be　achieved　simultaneously.