Aiming at the problems like short of samples and low accuracy in the fault diagnosis of planetary gearboxes,a fault diagnosis method that combined with empirical mode decomposition(EMD),kurtosis-ranking and BP neural network was proposed.By using this method,the original vibration signal was decomposed by EMD,then the kurtosis-ranking for the inherent mode function (IMF) obtained by decomposition was carried out;the corresponding IMF was adaptively selected according to the number of training samples,and the energy value of selected IMF was input as feature vector into the BP neural network to build the fault model and complete the model diagnosis success rate test.The results show that the distinguishing degree of fault features is significantly increased after kurtosis-ranking,indicating that by using the improved algorithm,higher diagnosis success rate can be achieved when there are few training samples or uneven training data,and the indentification of failture mode of the planetary gearboxes can be completed.