A Cross-Domain Online Diagnosis Framework Under Small Fault Sample

Abstract

Cross-domain fault diagnosis in industrial applications presents a major challenge, especially when only a limited number of fault samples are available. Existing data-driven and transfer learning (TL) methods often struggle with real-time diagnosis, insufficient generalization across domains, and the inability to adapt to continuously evolving fault conditions. To address these limitations, this article proposes a novel cross-domain online fault diagnosis framework (CDODF). The framework leverages the contrastive language-image pretraining (CLIP) model to extract robust, domain-invariant features from limited fault data. To further enable cross-domain adaptation without costly fine-tuning, a lightweight Adapter module is introduced, which incorporates few-shot learning and online adaptation to target-domain features. Moreover, CDODF supports a continuous learning strategy that dynamically updates the model using accumulated target-domain data, ensuring long-term adaptability and diagnostic accuracy. Experimental results across scenarios, including cross-working conditions, cross-device diagnosis, and emerging fault types, show that CDODF consistently outperforms existing deep learning (DL), TL, and few-shot methods.