Abstract <p>Accurately determining the health of critical components contributes to the efficient operation of industrial equipment and systems. Domain adaptation has emerged as a potent tool for cross-domain diagnosis, particularly in real-world scenarios that involve variations in the distributions of the utilized training and test data. However, the current domain adaptation methods are overly concerned with feature alignment while neglecting feature discriminability and the distinguishability of the intrinsic structure within the target domain. This results in the misclassification of target samples according to decision boundaries. In response to this issue, a double-level discriminative domain adaptation network (DL-DDAN) for cross-domain fault diagnosis is proposed. The DL-DDAN aligns domain-level features via adversarial training and designs a class-level discriminative module and a sample-level discriminative module. On the one hand, the class-level discriminative module not only achieves class-level alignment, but also promotes intra-class compactness and inter-class separation by pushing features belonging to the same class closer together and maintaining sufficient separation between the features of different classes. On the other hand, the sample-level discriminative module is applied to the target samples to mine their potential distinguishable information. The experimental results obtained on bearing and gearbox datasets, under various working conditions and measurement points, demonstrated the effectiveness and superiority of the DL-DDAN.</p> Graphical abstract <p>The diagnosis framework of DL-DDAN, including Health classifier loss, Domain discriminator loss, Class-level discriminative loss, and Sample-level discriminative loss.</p>