Domain generalization in rotating machinery fault diagnostics using deep neural networks

Abstract The past years have witnessed the successful development of intelligent machinery fault diagnostic methods. Besides the basic data-driven fault diagnosis tasks where the training and testing data are collected from the same distribution, the more practical cross-domain problems have also attracted much attention considering variations of machine operating conditions. In the literature, most existing studies generally assume the availability of testing data during model training, thus facilitating explicit domain adaptations. However, this assumption poses obstacles in the application on real-time cross-domain fault diagnosis, where the testing data can not be obtained in advance. This paper proposes a deep learning-based domain generalization method for machinery fault diagnosis. A domain augmentation method is adopted to expand the available dataset. Domain adversarial training is implemented, and generalized features can be learned from different domains, which hold in new working scenarios without assuming the availability of testing data. Distance metric learning is also used to further enhance model robustness in fault classifications. Through experiments on two rotating machinery datasets, the effectiveness of the proposed method is validated, which is promising in on-line cross-domain fault diagnosis tasks.