Abstract: To improve the precision of the ferromagnetic material hardness prediction method based on Barkhausen noise(BN) and automate the method, a group of features based on the first derivative and two derivatives of the auto regression (AR) spectrum of BN were proposed. A feature extraction algorithm was designed to unify the dimension of the frequency domain characteristics, and the group voting principle of random forest algorithm was improved to reduce noise interference and computational complexity. Through the hardness prediction experiments of two metals, the expected results were obtained. Compared with the mean square error of the classical time domain prediction method (229.8 and 298.7, respectively), the mean square error of the prediction method by using the feature and algorithm is only 60.3 and 81.3, respectively. The proposed feature and method have obvious accuracy and superiority.