Using　karman　filter　to　train　the　neural　network　can　overcome　the　drawbacks　of　BP　algorithm　such　as　falling　into　local　minima　and　slow　convergence.　However,　as　the　inverted　pendulum　model　is　a　strong　nonlinear,　unstable　system.,　we　should　first　linearize　the　model　system.　This　will　lead　to　huge　linearized　error.　Particle　filter　can　be　applied　to　the　status　estimation　of　any　nonlinear　and　non-Gaussian　system,　and　without　linerizing　the　system,　it　has　no　linearized　error.　In　this　paper,　we　found　the　physical　model,　the　filter　system　equation　and　the　observation　equation　of　the　inverted　pendulum　controller　and　use　particle　filter　to　estimate　the　neural　network　parameters.　We　compare　the　control　effect　between　karman　filter　and　particle　filter　in　the　mode　of　off-line.　The　simulation　results　show　that　the　performance　of　particle　filter　improves　markedly　than　karman　filter　both　on　speed　and　precision.