Aiming　at　the　problem　of　fault　feature　extraction　of　rolling　bearings　under　the　environment　of　strong　background　noise,　an　improved　singular　spectrum　decomposition　(SSD)　method　based　on　singular　value　difference　spectrum　is　proposed　to　detect　bearing　faults.　First,　in　order　to　avoid　the　empirical　selection　of　embedding　dimension　length　in　singular　value　decomposition,　a　novel　self-adaptive　signal　processing　method　called　SSD　is　used　to　handle　the　signal.　In　SSD　method,　a　new　track　matrix　is　built　and　the　window　function　length　is　adaptively　selected,　and　then　a　number　of　single　components　can　be　obtained　by　decomposing　the　nonlinear　and　non-stationary　signals.　However,　energy　leakage　will　appear　in　the　reconstruction　process　of　SSD.　To　solve　this　problem,　the　singular　value　difference　spectrum　method　is　used　to　improve　the　reconstruction　process,　which　can　improve　noise　reduction　capability　of　SSD　and　effectively　make　use　of　the　useful　information　as　well.　Finally,　in　order　to　extract　the　fault　characteristic　frequency,　the　envelope　demodulation　method　based　on　Hilbert　transform　is　used　to　analyze　the　component　which　contains　the　fault　information.　Simulation　and　experiment　signal　analysis　verify　the　effectiveness　of　the　proposed　method.　©　2019,　Nanjing　Univ.　of　Aeronautics　an　Astronautics.　All　right　reserved.