Abstract: To address the strong dependence on model parameters, limited adaptability to non-typical faults, and insufficient interpretability of deep learning-based methods in cable fault diagnosis, a novel cable fault diagnosis approach based on a multi-granularity Koopman modal graph-perception structure is proposed. First, approximate Koopman modes, amplitudes, and growth rates of system signals are extracted via dynamic mode decomposition(DMD) as shallow features to characterize the temporal dynamics of the cable fault diagnosis system. Then, a graph convolutional network(GCN) is constructed to extract deep features from node voltage and current signals under the power grid topology, enabling the modeling of spatial information and local dependencies. Finally, shallow Koopman modal features at both local and global scales are fused with the deep features learned by the GCN to form a composite representation that integrates physical interpretability with data-driven learning, thereby enhancing the model’s robustness and its capability to identify complex and non-typical faults. Simulation results on the IEEE 14-bus system demonstrate that the proposed method outperforms the comparative approaches in terms of diagnostic accuracy and stability. The proposed framework exhibits strong potential for embedded deployment and shows promising engineering applicability in online cable monitoring and intelligent operation and maintenance scenarios.

Key words: cable fault diagnosis, graph convolutional networks, Koopman operator, dynamic modal decomposition