Electromyographic (EMG) signals are the superposition of activities of multiple motor units (MUS). Therefore it is necessary to decompose the EMG signal in order to reveal the mechanisms pertaining to muscle and nerw control. Various techniques hare been devised with regards to EMG decomposition. A recently proposed method using wavelet analysis required manual selection of appropriate wavelet coefficients for action potential (AP) clustering. However the accuracy of this method depends heavily on the operators' ability to select suitable wavelet coefficients. To avoid this subjective ambiguilty we are proposing a new method ahich employs the principal component analysis on all wavelet coefficients to identify the distinguishable features of APs. The present method can decompose EMG automatically and unambiguously, from data input to clustering. Furthermore, our experimental results have shown that the decomposition accuracy was slightly higher than that of the conventional wavelet method.