Conventional intelligent diagnostic models are built on the foundation that the training data and testing data are recorded under the same operating conditions, which neglects the fact that the operating condition of the rotating machinery usually varies. The feature distribution of the recorded data under one set of operating conditions may be inconsistent with the feature distribution of the data recorded under different operating conditions. It is therefore easy to create a significant distribution discrepancy between training data and testing data. To address this issue, an unsupervised domain adaptation approach, based on a symmetric co-training framework, is proposed in this study. In the proposed symmetric co-training framework, a universal feature extractor and two individual classifiers are built as the main elements. The structures of the two classifiers are symmetric, and its parameters are updated in a co-training style. The parameters of the feature extractor and the two classifiers are continuously updated via an adversarial training process. The cosine similarity of the predictions from the two classifiers is introduced in order to guide the adversarial training process, which not only minimizes the distribution discrepancies between source domain data and the target domain data, but also pushes the feature subspaces for different healthy conditions away from the class boundaries. The application of the proposed method to two sets of experimental bearing fault data validates that the proposed method can successfully address the domain shift phenomenon between the recorded data under different operating conditions.