Closing the Simulation-to-Reality Gap for Digital Twin-Assisted Fault Diagnosis: Sim2real Knowledge Transfer with Contrastive Learning
25 Pages Posted: 18 Jan 2024
Abstract
Digital twin technology, crucial for fault diagnosis in smart manufacturing, provides accurate models of physical equipment and generates labeled synthetic data efficiently. Despite its benefits, a challenge arises from the distribution divergence between synthetic and real-world data, affecting model generalization. To address this, we propose the Contrastive Sim2Real Adaptation (CSRA) approach. CSRA leverages a pre-trained model on the labeled source data for self-supervised learning and knowledge transfer to the real-world unlabeled target data, effectively bridging the domain gap between the simulated and real-world environments. This method addresses training data privacy concerns and eliminates the need for source data when adapting pre-trained models to real-world scenarios. It benefits both providers of pre-trained models and clients using these models for customized applications. CSRA demonstrates superior performance in handling large domain gaps, improving model generalization in new environments, and narrowing the synthetic-to-real-world gap compared to standard cross-domain models in extensive experiments.
Keywords: Transfer Learning, contrastive learning, domain adaptation, digital twins, fault diagnosis, smart manufacturing
Suggested Citation: Suggested Citation