Tldiag: a federated learning-based DB-mixer for privacy-preserving and fault diagnosis in transmission lines

Abstract

As a critical component of smart grids where faults occur frequently, the rapid and accurate diagnosis of transmission line faults is essential for enhancing system reliability and response efficiency. However, existing centralized diagnostic methods face significant challenges, including data privacy concerns, limited feature representation capabilities, and class imbalance issues in real-world power grids. To address these issues, this paper proposes an adaptive diagnostic framework based on federated learning and a dual-branch mixer (TLDiag) to address the privacy-sensitive cross-domain fault diagnosis problem in dynamic grid environments. Specifically, a distributed federated learning architecture is designed to ensure data privacy by enabling localized model training through collaborative client–server interactions. Furthermore, a dual-branch mixer, i.e., DB-Mixer, guided by a dual-branch attention mechanism (DBAM), is developed to enhance feature representation by jointly modeling spatial and channel-wise information. Additionally, an adaptive dual-field loss (ADF loss) is introduced, incorporating dynamic task weighting and physical constraints to effectively mitigate class imbalance and improve diagnostic robustness. Extensive experiments conducted on the IEEE 5-bus system demonstrate that TLDiag achieves superior performance in both fault type classification (97.53%) and fault location identification (98.34%), while exhibiting stable convergence across varying client scales. Compared to baseline methods such as MDCNN and CNN-LSTM, TLDiag significantly outperforms in accuracy and robustness. By deeply integrating federated learning with the physical characteristics of power systems, this approach offers a high-accuracy, privacy-preserving solution for real-time fault diagnosis in smart grid scenarios.