Wasserstein distance-enhanced variable edge-weight graph convolutional networks for industrial fault diagnosis.

Modern industrial systems generate large volumes of high-dimensional and correlated sensor data, making the fault diagnosis increasingly challenging. Traditional methods often struggle to handle such non-Euclidean and nonlinear structures, and they typically fail to exploit the intrinsic topological or relational information embedded in the data. These limitations hinder their effectiveness in capturing complex inter-variable dependencies, which are critical for accurate fault identification. In contrast, Graph Neural Networks (GNNs) offer a promising framework to model such structural information. However, many existing GNN-based models overlook temporal correlations or suffer from high computational costs due to fully data-driven graph construction. To address these challenges, we propose a Wasserstein Distance Variable Edge-Weight Graph Convolutional Network (WVEGCN). This method integrates a mechanism-informed adjacency matrix specific to chemical processes and introduces adaptive edge-weight coefficients to improve robustness. We also design a feature extraction method based on Wasserstein distance to distinguish fault types more effectively and apply a novel feature selection strategy to enhance representation. A random forest classifier is used to improve stability in the final diagnosis. Experiments on two benchmark datasets (TE and TFF) demonstrate that our method significantly outperforms existing approaches in both accuracy and robustness, showing strong potential for real-world fault diagnosis applications.