High Speed Centrifugal Hydrogen Compressors are big and critical equipments which are widely used in chemical enterprises. It is very important to monitor their condition on-line and diagnose their failure. Based on the research of EMD (empirical mode decomposition) and correlation dimension and experimental simulation, the vibration signals of a High Speed Centrifugal Hydrogen Compressor’s main spindle are collected when working. Then the signals are decomposed by EMD, their correlation dimensions are calculated and taken as fault feature. Finally, using BP algorithm of a neural network in which there are 3 lays, a satisfied effect of fault diagnosis of a High Speed Centrifugal Hydrogen Compressor has been got. The experiment has confirmed that the method is advanced, reliable and practical. A new method is provided for High Speed Centrifugal Hydrogen Compressors’ fault diagnosis.