As　the　prerequisite　for　obtaining　the　optimal　demodulation　frequency　band　(ODFB),　the　frequency　band　division　model　must　be　delicately　constructed.　The　division　models　in　existing　methods　of　obtaining　ODFB　are　vulnerable　to　the　background　noise　and　irrelevant　components,　which　will　cause　an　unsuitable　demodulation　frequency　band.　To　improve　the　performance　of　frequency　band　division　model,　this　paper　proposes　a　fault　diagnosis　method　called　power　spectrum　screening　combination-gram　(Psscgram).　Firstly,　a　frequency　band　screening　combination　model　based　on　the　auto-regressive　(AR)　power　spectrum　is　established　to　obtain　the　frequency　band　group　of　vibration　signal.　Secondly,　the　negative　entropy　(NE)　is　used　to　quantify　the　fault　characteristics　of　the　frequency　band　group.　The　frequency　band　with　the　largest　NE　value　is　extracted　from　the　frequency　band　group,　whose　characteristic　frequency　is　calculated　via　envelope　spectrum.　Finally,　the　robustness　and　the　anti-interference　of　the　proposed　method　is　verified　by　simulated　signals　with　different　signal-to-noise　ratio.　Moreover,　the　ODFB　extraction　ability　of　the　proposed　method　is　proved　by　the　analyses　of　three　sets　of　bearing　experimental　signals.