@Article{sdhm.2026.088908,
AUTHOR = {Fanhao Zhou, Hui Wang, Wenqiang Liu, Haonan Yang, Hongrui Wang, Xinlong Liu, Zhigang Liu},
TITLE = {HPCLIP-AD: Hybrid Prompt Adaptation of CLIP for Anomaly Detection in Electrified Railway Catenary Support Components},
JOURNAL = {Structural Durability \& Health Monitoring},
VOLUME = {},
YEAR = {},
NUMBER = {},
PAGES = {{pages}},
URL = {http://www.techscience.com/sdhm/online/detail/28387},
ISSN = {1930-2991},
ABSTRACT = {Reliable inspection of catenary support components is challenging because defective samples are scarce, defect regions are often small, and outdoor railway scenes contain complex background interference. To address these challenges, this paper proposes a hybrid prompt adaptation framework based on Contrastive Language-Image Pre-training (CLIP), termed Hybrid Prompting CLIP for Anomaly Detection (HPCLIP-AD), for joint image-level defect identification and pixel-level defect localization. The framework combines static prompts learned from auxiliary industrial anomaly data with image-conditioned dynamic prompts to strengthen defect-related semantics, and introduces an anomaly feature enhancement module that groups patch embeddings into semantic clusters and selectively aggregates the most suspicious cluster, thereby suppressing irrelevant background responses. Experiments on a real-world catenary inspection dataset show that HPCLIP-AD achieves image-level area under the receiver operating characteristic curve (AUROC), average precision (AP), and F1 scores of 98.9%/97.7%/96.0% and pixel-level AUROC, AP, and F1 scores of 99.3%/72.4%/67.5%, respectively, outperforming representative anomaly detection baselines. These results indicate that the proposed framework can provide both reliable component-level defect alarms and localized defect cues for practical catenary inspection and maintenance.},
DOI = {10.32604/sdhm.2026.088908}
}