Objective Muscle fatigue is commonly associated with the musculoskeletal disorder problem. Surface electromyography (SEMG) provides important muscle activation information of exerted forces. This paper proposes a method to discriminate SEMG in maximum volunteer contraction (MVC) and fatigue state with the feature of certain frequency’s corresponding energy. Methods We recorded the SEMG signals on the right upper limbs from ten young men. The decomposition coefficients of each node on level 3 and level 4 for SEMG were gained by wavelet packet transform. The corresponding band energy of each node was normalized as feature vectors. The feature vectors were entered into back propagation neural network and support vector machine to recognize muscle fatigue. Results The muscle fatigue and MVC could be identified by four-level wavelet packet transform and back propagation neural network. The accuracy reached 87.5% with 7-fold cross-validation. Conclusions Recognition with wavelet packet energy transform can effectively distinguish muscle fatigue from different states of muscle contraction.