In regards to the short-term impact signal feature generated when the automaton works, first, the automaton movement patterns are decomposed, and vibration signals corresponding to the motion of the fault component are intercepted as the analysis object. Second, the multi-wavelet packet decomposition is used to intercept signals, and the frequency components and distribution of energy are studied. Signals after wavelet packet decomposition, which have larger band energy, are reconstructed to conduct local wave decomposition, while local wave singular spectrum entropy, marginal spectrum entropy and spatial spectral entropy are extracted as fault features to quantitatively describe the signal state changes in time domain, frequency domain and energy. Finally, the global optimization capability of genetic algorithms is used to optimize the parameters of support vector machine （SVM）. A GA-SVM model is established and the classification recognition results based on the extracted features are compared to the identify results of space exhaustive search SVM.