A Two-Stage and Hierarchical Contrastive Learning Method for Power Equipment Defect Detection
TL;DR
A two-stage framework that decouples component localization from operating-state recognition is proposed that consistently outperforms conventional multi-class detection methods, especially on rare defect categories, while effectively reducing false alarms in practical inspection scenarios.
Abstract
Power equipment defect detection is essential for ensuring reliable power grid operation. Existing power equipment defect detection methods typically formulate the task as multi-class object detection, but their performance is limited by long-tailed defect distributions and the neglect of normal operating conditions. To address these challenges, we propose a two-stage framework that decouples component localization from operating-state recognition. Specifically, a conventional detector first localizes equipment components, followed by a hierarchical contrastive learning framework for state recognition. A dual-memory bank is introduced to capture coarse-grained component semantics and fine-grained state prototypes, enabling discriminative feature learning through coarse-to-fine contrastive optimization. We further construct a real-world power equipment inspection dataset with hierarchical component–state annotations. Extensive experiments demonstrate that the proposed framework consistently outperforms conventional multi-class detection methods, especially on rare defect categories, while effectively reducing false alarms in practical inspection scenarios.