Closing the Simulation-to-Reality Gap for Fault Diagnosis in Unknown Environment: A Sim2Real Knowledge Transfer Approach With Contrastive Learning

Training a supervised model for fault diagnosis often requires labeled data samples, typically gathered from controlled environments because collecting such data in real-world scenarios is expensive and labor-intensive. As a result, these datasets often lack scalability and fail to capture the wide range of fault types encountered in practice. To address this limitation, synthetic data are frequently used to generate virtually unlimited labeled samples, providing a diverse array of fault patterns. However, despite its benefits, synthetic data introduces a challenge due to the distribution divergence between synthetic and real-world data, which can affect model generalization. Domain adaptation approaches has been used to mitigate the distribution divergence but it usually failed in such simulation-to-real-world cases as the large volume of labeled synthetic data dominates the feature extractor and make the knowledge hard to transfer to the unlabeled real-world data domain. To address this challenge, we propose the contrastive Sim2Real adaptation (CSRA) approach, which pretrains model on the labeled synthetic data and knowledge transfer to real-world unlabeled data in a self-supervised manner. By only using the pretrained model from the synthetic data, CSRA does not depend on the labeled synthetic data during the knowledge transfer; hence the feature extractor can focus on the real-world data. Then, CSRA employs contrastive learning techniques to align the feature distributions of synthetic and real-world data in a self-supervised way, thereby enhancing the robustness and accuracy of the fault diagnosis model. Our extensive experiments demonstrate that CSRA outperforms standard cross-domain models in handling large domain gaps. Specifically, CSRA improves model generalization in new environments, significantly narrowing the synthetic-to-real-world gap. The results indicate that our approach not only enhances the reliability of fault diagnosis systems but also provides a scalable solution for real-world applications, reducing the dependency on costly and labor-intensive data collection processes.