In　the　traditional　fault　diagnosis　technology,　classical　life　and　reliability　tests　require　sufficient　sample　size　when　diagnose　the　faults　and　forecast　the　future　states.　However,　there　is　even　less　sample　size　for　machinery　products,　especially　for　major　equipment.　The　Support　Vector　Machine　based　on　Statistical　Learning　Theory　can　solve　this　problem.　In　this　paper,　a　forecast　model　for　reactor　coolant　pump　which　combines　LSSVM　(Least　Squares　Support　Vector　Machine)　and　Time　Series　model　is　constructed.　We　studied　the　impact　to　forecast　accuracy　which　caused　by　embedding　dimension　M,　kernel　function　σ　and　regularization　parameter　γ.　Meanwhile,　the　performance　of　LSSVM　is　verified　by　simulation　data　and　field　data.　Then　LSSVM　is　used　to　predict　vibration　signals　of　reactor　coolant　pump.　As　it　is　certified　that　the　forecast　data　could　match　the　actual　data　preferably　and　has　achieved　good　results　in　forecasting　field　data.　©　2008　IEEE.