Ensemble Fault Identification Method for Converter Valve Based on Multi-Model Feature Fusion and Dynamic Graph Network Modeling

Abstract

Accurate fault identification is essential for reliable condition monitoring of converter valves in flexible DC transmission systems. Existing single-model methods are limited by architecture-specific feature extraction, while conventional multi-model fusion methods often neglect semantic inconsistency and sample-dependent relationships among heterogeneous representations. To address these limitations, this study proposes a multi-model fault identification framework based on cross-attention and dynamic graph convolutional networks (DGCNs). LSTM, CNN, Transformer, and TCN are employed in parallel to extract complementary features from multichannel waveform signals. Multi-head cross-attention is then used to align heterogeneous representations, which are subsequently treated as graph nodes for sample-dependent relational modeling and adaptive fusion using DGCN. Experiments are conducted on event-level independent electromagnetic-transient simulation data covering five operating conditions and ten fault types. The proposed method achieves an overall accuracy of 93.6% and a Macro-F1 of 0.904, outperforming the single-model and conventional static-fusion baselines. Under 15% Gaussian noise, the Macro-F1 remains 0.824. In addition, leave-one-operating-condition-out evaluation yields an average Macro-F1 of 0.875, demonstrating satisfactory generalization to unseen operating conditions. These results indicate that the proposed framework provides robust and adaptive fault identification for converter valves.