Abstract: Using the sound signals of opening and closing to realize the mechanical fault diagnosis of circuit breaker is easily affected by background noise. And the results of blind source separation algorithm used for sound signals' separation are disorderly. In this paper, a method in which blind source separation of sound signals is combined with the support vector data description with negative samples (NSVDD) is proposed for diagnosing the mechanical fault of circuit breaker. First, fast independent component analysis (FastICA) based on the maximization of negative entropy is used to realize the separation of the sound signals during the closing of circuit breaker. Then, the Gammatone frequency cepstrum coefficients (GFCC) of each separated signal are extracted according to the characteristics of human hearing. Meanwhile, variational mode decomposition (VMD) is performed on each separated signal to obtain its band-limited intrinsic mode functions (BLIMF) components to extract the singular spectrum entropy, energy entropy, and kurtosis entropy of the signal, then, the entropies are combined with the reduced dimensionality GFCC coefficients as combined feature vector of the sound signal. The single-value classification algorithm NSVDD is applied to identify the combined feature vector to eliminate the influence of noise. The experimental results show that the method based on blind source separation and NSVDD can be adopted to identify the mechanical fault of circuit breaker accurately under noisy background.