Considering the problem that stability of surface Electromyographic Signal (sEMG) based human-machine interface (HMI) gradually declines as fatigue takes place in muscles, we propose a novel method for updating samples to improve incremental online training algorithm for support vector machine (SVM). We study the changes of sEMG when muscle fatigue occurs using a method based on continuous wavelet transform, and then applies the improved incremental online SVM for sEMG classification. Experiment results show that the proposed algorithm can be used to improve the classification accuracy and training speed significantly. Furthermore, this method effectively diminish the influence of muscle fatigue during long-term operation of sEMG based HMI.