Federated Local Staircase Training for Photovoltaic Cell Fault Detection: Improving Rare Class Detection

作者：Zhefan Jin, Jiaojie Li, Renjie Zhang · 发表于：2025 4th International Conference on Energy and Electrical Power Systems (ICEEPS) · 年份：2025 · DOI：10.1109/ICEEPS66790.2025.11239705

To address the challenge of highly imbalanced rare class distributions and the difficulty of precise identification in photovoltaic (PV) cell fault detection, this paper proposes a novel approach that integrates federated learning with local staircase training. While preserving data privacy for all participants, the proposed method progressively unlocks training classes in stages, effectively enhancing the model's capability to detect rare fault types. Specifically, the system initially trains only on major classes, and as detection accuracy for previous classes improves, additional fault categories are gradually introduced, ultimately achieving joint optimization across all classes. Experimental results demonstrate that the proposed method achieves significantly higher mAP for rare classes on several real-world PV cell defect datasets, surpassing traditional federated learning strategies and exhibiting superior generalization and application potential.