Aiming at the non-stationary characteristics of wheel/rail fault noise signals, a wheel/rail fault diagnosis method based on Empirical Mode Decomposition(EMD) and neural network is put forward. Frist of all, wheel/rail noise signals are decomposed into several Intrinsic Mode Functions(IMF), then a number of IMFs including main fault information are selected. The energy and kurtosis features of these IMFs are extracted, and the kurtosis of these IMFs are integrated into a muti-scale kurtosis feature. Finally the energy feature of these IMFs and muti-scale kurtosis feature are served as input parameter of neural network to identify the fault pattern of wheel/rail system. The analysis results of wheel flats, rail wavy wear and normal state show that the approach of neural network diagnosis method based on EMD method extracting feature parameters has a higher fault recognition rate than that based on wavelet packet method. This method can classify and identify wheel/rail fault patterns accurately and effectively.