Due　to　the　existence　of　probe　pre-travel　error,　the　existing　researches　consider　single　or　double　influencing　factors　for　error　compensation.　However,　many　experimental　statistics　show　that,　due　to　the　anisotropy　of　touch　trigger　probe,　factors　such　as　signal　transmission　delay,　inspection　speed,　ball　radius,　rod　length,　probe　gravity　and　normal　vector　of　measuring　point　on　ball　surface　will　affect　the　trigger　time　of　inspection　signal,　resulting　in　probe　comprehensive　pre-travel　error,　so　it　is　difficult　to　compensate　accurately.　When　inputs　are　multiple　error　influencing　factors,　the　efficient　approximation　algorithm　of　BP　neural　network　is　helpful　to　solve　the　measurement　error　output　problem,　so　as　to　improve　on-machine　inspection　accuracy　effectively.　According　to　the　on-machine　inspection　principle　of　independently　developed　horizontal　gear　grinder　L300G,　taking　multiple　error　influencing　factors　as　the　input　node　and　probe　comprehensive　pre-travel　error　as　the　output　node,　then　the　prediction　model　of　probe　comprehensive　pre-travel　error　based　on　BP　neural　network　is　established.　After　error　compensation,　the　standard　sample　gear　on-machine　inspections　of　gear　grinder　were　carried　out.　Results　show　that,　before　and　after　error　compensation,　the　tooth　lead　accuracies　are　both　at　level　4;　after　error　compensation,　the　tooth　profile　accuracies　are　improved　by　2　grades　and　are　at　level　4,　which　is　consistent　with　that　of　Gleason　inspections.　Conclusion　verifies　the　correctness　of　model,　which　is　expected　to　be　popularized　in　the　high-precision　on-machine　inspection　system　of　domestic　low-cost　gear　grinder.　©　2022　Editorial　Office　of　Chinese　Journal　of　Mechanical　Engineering.　All　rights　reserved.