A fault diagnosis system of rotating machine is proposed in this thesis. The kernel of this diagnosis includes back propagation neural network and fuzzy reasoning engine. The experimental rotor system is produced by Bently Nevada Corporation. It is applied to simulate the specified rotating machine faults. This fault diagnosis system includes two stages: 1) pre-processing of vibration signals, 2) diagnosis of faults. In first stage, the principal axis of full spectrum with plane information of two sensors is used to emphasize the vibration phenomenon; it is more robust than traditional half spectrum. In addition, the fault signal sequence minus the average of standard signal’s sequence is applied to magnify the characteristic of fault signal in order to avoid the ambiguity of similar fault signal. The full spectrum and principal axis of full spectrum in pre-processing of vibration signals will be fed into the diagnosis system. In second stage, the training phase is processed before diagnosis phase. This diagnosis system is constructed with a set of individual neural networks, each network corresponds to a specified machine fault, and the evaluated indexes from each network are synthesized by a fuzzy reasoning engine in order to precisely diagnose the machine faults. Furthermore, the Taguchi’s method is applied to optimum the parameters of neural network and fuzzy reasoning engine in order to improve the efficiency of this diagnosis system. The advantage is that the trained epoch can be dramatically reduced, the over fitting problem can be avoided and the accuracy can be improved. Thus, the proposed methodology can improve the accuracy and robustness of diagnosis.