An adaptive and mechanism-informed dual-stream graph neural network for PEMFC fault diagnosis *

Abstract Accurate fault diagnosis of proton exchange membrane fuel cell (PEMFC) systems using multi-sensor data is crucial for safe operation. Conventional deep learning methods suffer from limitations, they cannot model complex non-Euclidean spatial correlations among sensors and ignore multi-physics coupling and multi-scale temporal dynamics, hindering effective spatio-temporal feature fusion. To address these issues, a dual-stream spatio-temporal dynamic graph convolutional network (DS-STDGCN) is proposed. This dual-stream framework combines adaptive data-driven learning with a mechanism-informed structural prior. The adaptive data-driven stream learns sample-dependent sensor associations, while the mechanism-informed prior stream uses relationships predefined from PEMFC mechanisms and variable definitions, weighted by transfer entropy estimated from training data. Chebyshev graph convolution extracts features on this weighted prior graph. Both streams employ multi-scale gated temporal convolutions to capture cross-frequency fault features and enhance sensitivity to early subtle faults. A node-wise gated fusion module adaptively balances the two streams. Experiments on real PEMFC datasets demonstrate that DS-STDGCN maintains high robustness and stability under varying operating conditions, high noise, and domain shifts, significantly reducing fault misdiagnosis and missed diagnosis rates.