Stockbridge dampers, critical components of transmission line integrity management, require precise defect detection to ensure grid reliability. While deep learning has emerged as powerful tools for identifying damper defects amidst complex environmental interferences and variable target morphologies, current approaches lack integrated privacy preservation mechanisms– a critical limitation given the fragmented distribution of inspection data across regional utilities, which exacerbates data silos and impedes collaborative model refinement. This study introduces a privacy-aware federated learning framework synergizing an optimized YOLOv11 architecture with systematic privacy-preserving mechanisms for damper defect diagnostics. Our methodology fundamentally redefines data governance by implementing localized client training with the Federated Averaging algorithm (FedAvg) for secure multi-party parameter aggregation, thereby eliminating raw data transmission while ensuring model convergence. Three pivotal contributions distinguish this work. First, we establish the FDRD benchmark dataset comprising real-world transmission line inspection imagery across multiple defect scenarios, creating the first standardized evaluation dataset for damper condition analysis. Second, we develop a federated learning architecture integrating encrypted parameter exchange protocols that jointly address data privacy constraints and regional data fragmentation, enabling collaborative model enhancement without raw data centralization. Third, extensive evaluations demonstrate significant performance improvements over baseline models (YOLOv9/YOLOv10), achieving state-of-the-art metrics including 0.9 mAP50, 0.928 precision, and 0.785 recall while preserving detection robustness comparable to centralized training paradigms. We share our code at https://github.com/yd479/Fed-StockbridgeDefect.git.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号