In order to improve the analysis accuracy of Barkhausen signal, we first analyze the shortcomings of the traditional feature extraction method of Barkhausen signal. Based on the analysis of HilbertHuang transform theory, a new Barkhausen signal feature extraction method is proposed. The method includes both time and frequency information, theoretically has a very high analytical accuracy. By comparing the new feature with the traditional feature, it is found that the recognition rate of the new feature for classification is much higher than the traditional one, and the training sample size is smaller than the traditional one. In this paper, we also find that the recognition rate of a single traditional eigenvalue is very low, and the traditional eigenvalues can be combined to improve the recognition rate.