Pump fault diagnosis for vanadium redox flow batteries using an optimized smoothing-based graph convolutional network

Vanadium redox flow batteries (VRBs) are promising for large-scale energy storage, yet reliable pump fault diagnosis remains challenging due to complex multivariate coupling and limited labeled data. Graph convolutional networks (GCNs) provide an effective way to model variable relationships, but their performance is fundamentally limited by the oversmoothing problem, which leads to loss of feature discriminability in deep architectures. To address this issue, an optimized smoothing-based GCN (OS-GCN) framework is proposed. A smoothing factor (SF) is introduced to regulate feature propagation, and particle swarm optimization is employed to determine its optimal value, enabling a balance between information aggregation and feature preservation. Data association graphs for the GCN are constructed based on Euclidean distance (ED) relationships to capture system dynamics. Pump fault data are collected under both controlled and real-world conditions, providing the basis for an in-depth analysis of operating temperature effects, robustness to noise, and the applicability of the proposed model. Experimental results show that OS-GCN significantly outperforms existing fault diagnosis methods under controlled and real-world conditions. These findings confirm that the proposed framework effectively mitigates oversmoothing and improves the accuracy and robustness of VRB pump fault diagnosis.