In　order　to　solve　the　difficulty　of　features　extraction　of　compound　faults　in　underdetermined　state,　this　research　proposes　an　approach　to　extract　signal　features　by　combining　adaptive　generalized　S　transform　(GST)　and　non-negative　matrix　factorization　algorithm　(NMF).　The　adaptive　function　(AF)　is　introduced　to　optimize　GST.　The　optimized　GST　is　used　to　process　monitored　signals　to　get　the　time-frequency　features　matrix.　The　NMF　is　improved　by　Itakura-Saito　(IS)　divergence.　And　the　dimensionality　of　the　signal　time-frequency　matrix　is　reduced　by　it.　After　iterative　updating,　several　low-dimensional　matrices　are　obtained.　The　time-domain　waveforms　of　low-dimensional　matrices　are　reconstructed,　and　the　envelope　spectrum　analysis　is　performed　to　realize　compound　faults　diagnosis.　The　simulation　test　and　the　actual　bearing　compound　fault　signals　experiment　prove　that　this　method　can　effectively　extract　compound　fault　features　in　underdetermined　state　and　realize　bearing　compound　faults　diagnosis.