T-Guardian: A vision–language framework for telescope performance monitoring with anomaly-aware description generation

Large-aperture telescopes are essential facilities for astronomical data acquisition. Their system performance directly affects data quality and the reliability of scientific outcomes, making interpretable performance monitoring and human-readable status reporting critically important. However, existing data-driven and rule-based methods remain limited in simultaneously providing human-readable outputs and clear interpretability, while rule-based approaches rely heavily on manual rule construction. To address this issue, we propose T-Guardian, a vision–language framework for telescope performance monitoring that bridges the semantic gap between telescope images and status descriptions. The framework incorporates a Multi-scale Contrastive U-Block for cross-scale visual-text alignment, a residual attention path to preserve fine-grained visual details, and an Adapter module to map multimodal features into the language generation space. Experimental results show that T-Guardian outperforms traditional methods and general-purpose vision-language models, achieving a macro-F1 of 0.9857 for state classification and a BERTScore of 0.9343 for description generation. Moreover, expert evaluations by astronomical technology specialists rank T-Guardian highest among all evaluated models, further validating its effectiveness in producing reliable, interpretable, and human-oriented telescope monitoring reports.