This article takes the high-speed train axle box bearing as the research object. A bearing fault diagnosis method is proposed to deal with finite variable working conditions, which is based on the supervised auto encoder feature representation transfer. The feature sequences of different working conditions are mapped to the reference condition feature sequences. In this way, the influence of condition change on bearing fault feature is decreased. The migrated features are inputted into the fault diagnosis model based on the convolution neural network, which is pre-trained by the reference condition training feature sets. Then, the axle box bearing fault diagnosis is achieved under variable working conditions. The open bearing data of Case Western Reserve University and the high-speed axle box bearing data are utilized. Experimental results show that the accuracy of fault identification has been greatly improved after feature migration. The method can achieve the feature migration under different working conditions and reduce the distortion of fault features caused by the change of working conditions.