Aiming at  the problem of low accuracy in bearing fault diagnosis caused by insufficient typical samples of fault data in the industrial domain and agricultural machinery,  a bearing fault classification method based on improved adversarial distillation is proposed.  The adversarial knowledge distillation method is used to classify the bearing faults. Based on the soft labels of the teacher network, the student network produces samples similar to those output by the student network. The modification of the parameters of the student network proceeds as the discriminator evaluates the samples. In this paper,  an annealed modified adversarial distillation method is proposed to improve the robustness and generalization ability of the student network. With dynamic temperature training in adversarial distillation, the difficulty of generating samples is increased for more efficient utilization of information from the teacher network. The effectiveness of the method is verified through experiments based on the bearing fault dataset from Case Western Reserve University in the United States. The student network trained with the proposed method achieves an accuracy of 91.85% in the simulation of onsite bearing fault diagnosis classification task, with only 214602 parameters involved in the computation, which not only improves the accuracy of fault diagnosis but also saves computing resources of the equipment.