UniFault: A Fault Diagnosis Foundation Model from Bearing Data

Machine fault diagnosis (FD) plays a vital role in predictive maintenance by enabling early detection of potential failures and reducing unexpected downtime. However, current FD models are often tailored to specific operating conditions, limiting their ability to generalize across heterogeneous datasets. While foundation models (FMs) in vision and language have shown impressive generalization and adaptability through few-shot and zero-shot learning, extending these strengths to FD remains challenging. Unlike the large, uniform datasets in visual or textual domains, FD data are typically small-scale and highly diverse, with differences in sampling rates, channel numbers, and system configurations. This variability complicates the creation of a unified model that can robustly extract and learn discriminative features from such heterogeneous sources. To address these challenges, we propose UniFault, a foundation model specifically designed for fault diagnosis. UniFault introduces a unified data harmonization framework consisting of two core components: (1) a unification scheme that standardizes multivariate signals into structured univariate representations while preserving inter-channel dependencies, and (2) a cross-domain temporal fusion mechanism that alleviates distribution discrepancies and enhances data diversity for improved generalization. Pretrained on over 9 billion data points from a wide range of FD datasets, UniFault achieves superior few-shot adaptability and strong generalization. Extensive experiments on real-world benchmarks confirm that UniFault surpasses state-of-the-art methods, establishing a new foundation for scalable and resilient predictive maintenance.