The　early　impulse　fault　diagnosis　of　the　gearbox　in　rolling　mills　is　often　difficult　and　labour　intensive　because　the　gearbox　of　that　high　speed　machine　is　multi-shafting　transmission　system,　in　which　many　gearsets　and　rolling　bears　work　together　at　the　same　time　and　there　are　much　complex　frequency　structure　and　various　disturb.　A　new　time-frequency　method　based　on　the　wavelet　packets　technique　was　developed　and　used　to　extract　the　impact　feature　from　signals　collected　from　faulty　data　of　one　rolling　mills　gearbox.　The　method　improves　the　signal　to　noise　ration　so　that　results　obtained　using　this　method　represents　features　with　fine　resolution　in　both　low-frequency　and　the　high　frequency　bands.　The　results　of　analysis　indicate　the　validity　and　the　practicability　of　the　method　proposed　here.