Traditional　independent　component　analysis　may　produce　different　results　in　the　repeated　calculations　for　their　statistical　characteristics,　and　thus　they　are　unreliable　and　unstable.　This　paper　introduces　a　novel　ICA　algorithm　which　enhances　the　separation　performance　and　stability　by　clustering　evaluation.　Firstly,　the　improved　ICA　algorithm　runs　a　single　ICA　algorithm　for　several　times　with　the　different　initial　parameters;　secondly,　the　separated　components　produced　are　clustered　according　to　their　relevance;　lastly,　the　best　separated　components　are　selected　as　the　optimal　results　by　clustering　evaluation.　The　effectiveness　of　the　improved　ICA　algorithm　is　validated　in　the　simulation　by　typical　mechanical　signals.　The　proposed　method　is　also　applied　to　extract　the　effective　information　of　observed　signals　on　the　bulkhead　of　a　ship,　and　the　results　show　that　most　of　the　important　information　is　well　extracted.　This　research　provides　a　novel　approach　for　vibration　reduction　and　control　of　ships.