Support　Vector　Machines　(SVM)　is　one　of　most　important　algorithm　in　machine　learning　area.　The　choice　of　kernel　function　can　have　great　influence　on　classification　and　approximation　ability.　Choosing　appropriate　kernel　function　and　weight　parameters　is　one　of　the　keys　to　utilize　SVM.　Single　kernel　function　always　has　its　limitation　in　the　application.　We　propose　a　new　kernel　function　based　on　the　analysis　about　the　constitute　conditions　of　the　kernel　function　and　the　characteristics　of　different　kinds　of　kernel　function-linear　compound　kernel　function,　this　function　not　only　can　reduce　the　amount　of　parameters　of　the　kernel　function,　but　also　has　good　learning　ability　and　generalizing　ability.　And　we　have　tested　the　effectiveness　of　the　kernel　function　through　simulation.