Agent-Based Contrastive Domain Generalization for Fault Diagnosis

Data-driven fault diagnosis plays a core role in intelligent operations and maintenance in modern industries. In practice, the deployment of fault diagnosis models often takes place in highly dynamic working environments. Consequently, models trained on a specific working environment may not generalize well to data from different working environments (i.e., different distributions). The naive approach of training a new model for each new working environment is impractical. Contrastive learning emerges as a promising solution, aiming to learn domain-invariant embeddings by leveraging rich semantic relationships among samples from different domains. A simple strategy involves pulling positive sample pairs from different regions closer together while pushing negative sample pairs further apart. However, aligning positive pairs across different domains often hinders model generalization due to significant distribution discrepancies. To address this challenge, we propose an Agent-based Contrastive Domain Generalization (ACDG) approach for diagnosing rotating machinery faults. Specifically, our method replaces the original sample-level relationships with agent-level relationships, effectively alleviating the issue of positive alignment. Experimental results demonstrate that our proposed ACDG method outperforms state-of-the-art approaches on four real-world fault diagnosis datasets. The favorable performance of ACDG on new unseen target domains contributes to more practical data-driven methods operating in challenging real-world environments.