In　mechanical　fault　diagnosis,　the　traditional　blind　source　separation　algorithm　requires　the　number　of　observed　signals　which　are　more　than　the　number　of　source　signals.　However,　in　most　cases,　mechanical　failure　does　not　satisfy　the　conditions　which　generated　underdetermined　blind　source　separation　problem.　We　proposed　an　underdetermined　blind　source　separation　method　based　on　wavelet　packet　First,　the　wavelet　packet　analysis　was　used　to　get　a　signal　with　noise　reduction.　Then,　the　processed　signal　was　decomposed　and　reconstructed　by　wavelet　packet　analysis　to　get　a　series　of　reconstructed　signals.　Second,　calculate　the　kurtosis　value　of　the　reconstructed　signals　and　select　the　larger　one.　Finally,　the　new　set　of　observed　signal　consisting　of　the　reconstructed　signal　and　the　source　signal　was　analyzed　by　the　blind　source　separation.　Simulation　and　measured　results　show　the　effectiveness　and　feasibility　of　this　method.　Compare　the　different　analysis　results　of　underdetermined　blind　source　separation　method　based　on　continuous　wavelet　transform　and　EMD.　It　is　proved　that　the　algorithm　we　proposed　is　superior　to　the　other　two　methods　both　in　simulated,　measured　and　engineering　signals.