Abstract: In the traditional rolling bearing fault diagnosis, vibration characteristics are easy to be interfered by redundant noise, and fault characteristics can not be accurately classified, a rolling bearing fault diagnosis method based on wavelet denoising, EMD-AR spectrum analysis and ELM (extreme learning Machine) was proposed. The vibration signals of rolling bearings are firstly de-noised by wavelet threshold, then the de-noised one-dimensional signals are decomposed by EMD and the first six IMF components are extracted. The AR spectra of the first six IMF components are accumulated to obtain the EMD-AR spectra of the de-noised vibration signals, from which different fault conditions of bearings can be seen as a prior diagnosis. Finally, six eigenvalues of the denoised signal were extracted as samples. In order to avoid the contingency of the experiment, the ELM classification diagnosis model based on K-fold cross-validation was established. The diagnosis results show that the method can clearly classify the bearing fault situation, the classification accuracy is up to 100%, and can provide a new method for the bearing fault diagnosis.