The structure of the planetary gearbox is very complicated. When a fault occurs， its vibration signal appears non-linear and non-stationary feature， and the fault signal is weak. In order to accurately extract features expressing planetary gear failure information， the signal processing method of local mean decomposition （LMD） combined with S-transform is proposed， and transform it into time-frequency distribution image. First， the vibration signal is processed by LMD-S transform， then the interference is filtered by correlation analysis method and transformed into time-frequency distributed image. Subsequently， the non-uniform local binary pattern （LBP） is used to extract image texture features under different working conditions. Finally， the limited learning machine is used to identify three fault types. The accuracy of fault recognition reaches 90%， which proves the effectiveness of this method.