A new fault diagnosis method based on Volterra series and KPCA is proposed. In the proposed method, the Volterra series of four states, i.e. normal, rotor crack, rotor rub and pedestal looseness, are identified by the particle swarm optimization (QPSO) algorithm, then the Volterra kernel function are used as feature vectors to input into kernel principal component analysis (KPCA) for training and recognition. The experiment result shows the proposed method is very effective. The higher order Volterra kernels such as second-order, third-order kernel are used to recognize each fault when the fault is hardly distinguished by first-order Volterra kernel (linear kernel function).

A blind separation method of rotor’s fault sources based on variational Bayesian independent component analysis (VBICA) is proposed. This method can directly separate the signals of mechanical sources in noisy environment. In this method, the unknown noise need not to be regarded as an independent source, and the denoising preprocessing is not necessary either. Then, this method is compared with the traditional blind source separation method for machine faults. Finally, this method is applied for the fault sources separation of rotor system. Experiment results show that this method is very effective.

The definition and algorithm of fractional cepstrum are introduced. A new machine fault diagnosis method based on fractional cepstrum is proposed. The experimental result shows that fractional cepstrum is superior to the traditional cepstrum. It can suppress the interference and enhance the quality of the diagram spectrum.