Aiming　at　the　disadvantages　of　the　traditional　K-means　clustering　algorithm,　a　new　algorithm　based　on　density　is　proposed　to　remove　the　noises　and　outliers　in　this　paper.　This　algorithm　determines　whether　a　point　is　a　noise　or　not　according　to　the　density　of　the　point.　Experiments　show　that　this　algorithm　can　effectively　eliminate　the　influence　of　the　noises　when　the　K-means　algorithm　searches　cluster　centers　in　the　samples.　Then　the　subtractive　clustering　algorithm　is　used　to　initialize　the　clustering　centers　of　the　K-means　algorithm,　meanwhile　the　number　of　cluster　centers　is　gotten.　The　improved　K-means　algorithm　is　taken　to　optimize　the　structure　of　RBF　neural　network,　and　the　results　of　experiments　on　the　typical　function　approximation　show　that　the　proposed　algorithm　has　the　better　approximation　ability.　©　2016　IEEE.