The reliable operation of power grids is increasingly dependent on advanced fault diagnosis systems capable of identifying both known and emerging fault types. Traditional and many contemporary methods often struggle in open-set environments where new fault types may arise, posing a significant challenge to maintaining grid stability. In this study, we propose an advanced open-set fault diagnosis approach for power grids, utilizing a novel text-image matching technique. Our method integrates a fine-tuned BERT model for extracting domain-specific textual features with a Swin Transformer-based model for capturing multi-scale visual features. These features are fused using a contrastive learning framework, enabling the system to accurately diagnose faults based on textual descriptions, even when those faults have not been previously encountered in the visual domain. This approach not only improves the detection of known faults but also provides robust diagnostic capabilities for new and unforeseen fault types, offering a powerful tool for enhancing the resilience and reliability of modern power grids.