As the characteristics of presents the fault features of vibration signals are hard to extract due to its nonlinearity, no-stationary and aperiodicity.In the light of the low fault recognition rate in an automata machine for this reason, a new approach is proposed which combines multiple scale sample entropy and variable predictive model based on the class discriminate (VPMCD). First, wavelet threshold noise reduction is applied to the collected signals. Then, the vibration signals are decomposed into several signals in different scales, the sample entropy of different signals are calculated and the sensitivity to fault characteristics are selected to constitute the fault features. Finally, the VPMCD is establishedbased on the fault features to recognize and classify the automata faults. The experimental results show that this method can accurately distinguish several typical fault forms of automata.