Abstract: Considering the problem of hardly determinate input signals in machine fault diagnosis method based on the system identification, and the characteristics of the abundant informa-tion, non-stationary, poor repeatability and reproducibility in the operating process of the mechanical equipment, here, combined blind identification of nonlinear time model and factorial hid-den Markov Model (FHMM), a fault diagnosis approach named as BSI-FHMM, is proposed. This approach is that blind identi-fication of nonlinear time series model is used as a feature ex-traction, and FHMM as a classifier, the proposed approach has been successfully completed in the speed-up and speed-down process of rotating machinery. At the same time, this approach is compared with another two fault diagnosis approaches named FFT-FHMM, wavelet-FHMM respectively. In the FFT-FHMM and wavelet-FHMM recognition approaches, the Fourier trans-formation and wavelet transformation is used as a feature ex-traction respectively, FHMM as a classifier. Experiment results show that the BSI-FHMM recognition approach is supe- rior to the FFT-FHMM and wavelet-FHMM recognition approaches.

Key words: Blind system identification (BSI), Factorial hidden Markov model (FHMM), Fault diagnosis, Nonlinear time series, Pattern recognition