one kind of fusion wavelet characteristic and the generalized distinction analysis （GDA） extraction method was proposed and it was applied to descript the efficiency of dimension reduction for multi-scale feature of SEMG. First, wavelet decompose was employed on SEMG, and the mean of absolute value for wavelet coefficient at each level was extracted as original feature vector, then the dimension by GDA was reduced, the new feature of lower dimension was obtained, which was used for classification by Bayes classifier. By selecting felicitous levels of wavelet decomposition as well as pretty kernel parameter and new dimension, it can result in good performance in the pattern recognition of six movements including wrist inward，wrist outward, wrist up, wrist down, fist clench and fist stretch with the accuracy of above 97 % for three testers. The research indicated that the proposed method can successfully acquire the main component of multi-scale SEMG and its feature.