This　article　proposes　a　classifier-based　approximator　to　compensate　for　friction　in　a　high　accelerated　positioning　system.　Since　friction　function　is　globally　non-smooth,　an　unsupervised　k-means　clustering　algorithm　is　adopted　to　classify　the　friction　into　micro　and　macro　motion　segment,　then　the　frictions　are　estimated　with　two　sub-approximator　in　the　corresponding　segment　respectively.　Due　to　the　unsupervised　classification　of　friction,　the　classifier-based　approximator　can　realize　universal　approximation　of　nonlinear　friction　with　high　precision.　Finally,　comparative　experiments　on　a　high　accelerated　position　system　driven　by　voice　coil　motors　are　conducted　to　verify　the　effectiveness　of　the　proposed　method.　The　proposed　method　can　reduce　the　root　mean　square　error　(RMSE)　of　tracking　by　57.2%,　19.1%,　and　27.4%　compared　with　a　parametric　model,　recurrent　neural　network　(RNN),　and　incremental　extreme　learning　machine　(I-ELM),　respectively.