Domain Transferability-Based Deep Domain Generalization Method Towards Actual Fault Diagnosis Scenarios

Recent years have witnessed the successful development of using knowledge transfer strategies to tackle cross-domain fault diagnosis problems. Most existing studies generally assume the availability of target data during training to guide the domain adaptation process. Actually, this assumption is difficult to hold in many real engineering scenarios, where fault data cannot be obtained in advance. Therefore, how to mine universal and effective diagnosis knowledge only from the source domain and generalize it to unseen target tasks is crucial for further developing intelligent cross-domain fault diagnosis. With this in mind, this article proposes a novel domain transferability-based deep domain generalization (DT-DDG) method. To achieve excellent diagnostic performance on the unseen domain, an adversarial training scheme combining the dynamic weighting strategy and a batch spectral penalization regularization term is elaborated so that the domain-invariant and discriminative features can be extracted from multiple source domains with potential discrepancies, which can hold in new target domain without assuming the available of fault data. Moreover, the network architecture is further optimized to address the category shift problem, thus taking full advantage of more source domains and learn more abundant diagnosis knowledge. On multiple generalization tasks using five datasets, the effectiveness and superiority of the DT-DDG method are validated, and the results show its great promise in actual fault diagnosis scenarios.