Domain adaptation has significantly promoted the development of intelligent fault diagnosis under variable working conditions. However, previous work developed the transferrable model based on source label information and cross-domain distribution discrepancy, while neglecting the intrinsic correlation of samples. This results in a decrease in transferability. To avoid this drawback, a self-supervised adversarial network (SSAN) is proposed in this paper, which explores sample similarity to guide feature alignment. Experiments on the gearbox dataset demonstrate the promising performance of SSAN in cross-domain fault diagnosis.