Knowledge-informed deep networks for robust fault diagnosis of rolling bearings

•Propose a knowledge- informed deep network (KIDN) framework for the health diagnostics of bearing systems.•Propose a generalizability-based adaptive network design strategy to improve the robustness and stability of the KIDNs.•Four bearing systems are utilized to demonstrate the effectiveness of the proposed approach. Effective fault defection is of critical importance in condition-based maintenance to improve the reliability of engineered systems and reduce operational cost. This paper introduces a knowledge-informed deep learning approach to fuse prior knowledge and critical health information extracted from raw monitoring data for robust fault diagnosis of rolling bearings. A set of knowledge-based features is first extracted based on prior knowledge of engineered systems. A knowledge-informed deep network (KIDN) is then designed to leverage these knowledge-based features with data-driven machine learning for the accurate prediction of bearing faults. To further enhance the generalizability of deep networks for fault diagnosis and alleviate extensive tuning efforts, a novel generalizability-based adaptive network design strategy is developed based on constrained Gaussian process (CGP) to quickly obtain the promising architectures for the development of knowledge-informed deep networks. Specifically, it involves the training of a constrained Gaussian process (CGP) surrogate model to predict the generalizability of KIDN and seeking potential improvements by exploring alternative network architectures within a vast design space. Four experimental case studies are implemented to validate the proposed methodology.