Double-classifier adversarial learning for fault diagnosis of rotating machinery considering cross domains

Deep learning methods have been demonstrated remarkable success in machine fault diagnosis under the constraint of identical distribution between training datasets and test datasets. However, achieving such conditions in practical scenarios remains challenging. Variations in working conditions lead to distinct distributions in fault data, while acquiring sufficient labeled fault data is often difficult. To address these problems, a double-classifiers adversarial learning network (DC-net) method is proposed. Firstly, a specialized network structure is designed, containing two classifiers, which align the source and target domains through the utilization of an adversarial training strategy. Secondly, conditional entropy and locally Lipschitz term are integrated into the loss function to force decision boundaries away from data-dense areas, precisely classifying different fault modes. State-of-the-art results are achieved across four cases, with test accuracy exceeding 80% in most instances. Notably, in single-source bearing fault diagnosis, the average test accuracy reaches 98.89%. These experimental results reveal the reliability and generalizability of the constructed model.