<p>Domain adaptation is a branch of transfer learning widely used in engine cross-domain fault diagnosis. However, most studies aim at minimizing the distance between source and target domains, ignoring that the distance between domains is too close to reduce the feature discrimination between classes. To solve these problems, a turn-based domain adversarial network for cross-domain fault diagnosis is proposed herein. Based on model transfer initialization and adaptive loss function weight technology, a coding network suitable for the original vibration signal is constructed. Then, the turn-based adversarial method is used to weaken the intense confrontation in a single training cycle to improve the feature recognition ability and adversarial stability. In the cross-fault degree and cross-speed domain tasks, the algorithm was compared with the traditional domain adaptation method. The experimental results indicate that the average accuracy of the proposed method on the engine dataset is 86.91%, which is higher than that of other methods by more than 6.90 %.</p>