To solve the ambient interference in high voltage circuit breaker online diagnostics which is based on acoustic signal analysis, a blind source separation method is proposed for sound disturbance signal recognition. A improved potential function is used to estimate sources, and then ensemble empirical mode decomposition (EEMD) is applied into calculating several intrinsic mode functions (IMF), multidimensional signals are reconstructed in line with the number of sources. The sound signal blind source separation is completed by Fast Independent Component Analysis (FastICA) which is optimized by a quasi-Newton Method. The sound signal component of breaker closing state identification is obtained according envelope features comparison. The results show that, the proposed method can make the useful acoustic signal generated by the circuit breakers’ action extracted effectively from mixed voice signal while the information of sources is unknown.