The vibration signals of the body and bogie of a high speed train are always affected by the bogie′s mechanical faults, which can be predicted using the nonlinear features extracted from the monitoring data. A new feature-analysis method is proposed based on ensemble empirical mode decomposition (EEMD) and permutation entropy. First, fault vibration signals are decomposed into a series of narrow band intrinsic mode functions (IMFs) using EEMD. Then, the permutation entropy of these IMFs and initial signals are calculated as multi-scale complexity measure feature vectors. Finally, the feature vectors are transformed in a least-squares support vector machine to classify and identify the operating conditions. Simulation and experimental results show that the recognition rate is above 95% under a running speed of 200 km/h, which proves the feasibility of this mechanical fault diagnosis method.