UDA-ROT: A cross-domain fault diagnosis method for large-scale systems

Abstract

Large-scale process industries frequently operate under multiple working conditions. These conditions share similar process technologies, operational principles, and failure mechanisms. Fault diagnosis for different working conditions can Fault diagnosis across these different conditions can be achieved by using programs with a unified label space for both source and target domains. Universal Domain Adaptation based on Reweighted Optimal Transport (UDA-ROT) is proposed for cross-domain fault diagnosis tasks. Firstly, the optimal transport theory is employed to characterize the differences between the source and target domains, in order to identify the unknown classes of the target domain and the private classes of the source domain. The inter-domain adaptation of public classes is achieved through a domain adversarial network. Secondly, a reweighting mechanism is devised relying on the domain adversarial output, in order to refine the optimal transport cost matrix and ensure more precise transport alignment. Finally, the global and local features of the target domain are fully explored to learn the optimal transport problem from the target samples to their prototypes, in order to encourage interclass separability and intraclass consistency of the target clusters. The proposed method provides an effective solution for large-scale systems fault diagnosis.