Abstract: A turbine vibration faults diagnosis method by using Gaussian Mixture Models was proposed. The original turbine vibration faults signal is decomposed and reconstructed by wavelet packet analysis method, which act as a filter. Then the character of the vibration signal is picked up and used to set up the GMM. For each fault situation, taking its several set of the fault data as training data, an identifying cell for this fault situation is created. The maximum likelihood estimation of parameter of identifying cell is solved with EM algorithm. At last, the unidentified data is input to every identifying cell, and the maximum probability cell is found out, and the fault of this cell is the last diagnosis result.

Key words: Gaussian Mixture Model (GMM), faults diagnosis, wavelet packet analysis, EM algorithm