Abstract: In the era of big data, intelligent fault diagnosis is an important tool to guarantee the healthy operation of the machinery. In order to accurately identify the health states of the machinery, it is necessary for intelligent fault diagnosis to collect massive available data and then train an intelligent diagnosis model. In real cases, however, there is no sufficient available data to train a fault diagnosis model with high diagnosis accuracy, which limits the engineering application of intelligent fault diagnosis. Fortunately, the data collected from the laboratory machines have sufficient typical fault information and label information. These data also have related fault information to the collected data from the real-case machines. Therefore, a deep transfer diagnosis method is proposed to identify the health states of the real-case machines by using the fault diagnosis knowledge from the laboratory machines. In the proposed method, the domain-shared deep residual networks are first used to extract transferable features of the collected data from different machines. Then, the regularization terms of domain adaptation are introduced into the training process of the deep residual networks so as to establish a deep transfer diagnosis model. The proposed method is verified by a transfer diagnosis case from the laboratory bearings to the locomotive bearings. The results show that the proposed method is able to use the fault diagnosis knowledge from the laboratory bearings to identify the health states of the locomotive bearings.