MS-DDGformer: A multi-scale synergized dual-driven graph transformer for industrial process fault diagnosis

Fault diagnosis in modern industrial systems is challenged by complex failures that require sophisticated spatio-temporal modeling. Although Graph Transformers (GTs) show promise, existing models often lack effective architectural coupling or are tuned for non-industrial tasks. They struggle with nonstationary process dynamics and fail to jointly model temporal dependencies and static topology. To overcome these issues, a Multi-Scale Synergized Dual-Driven Graph Transformer (MS-DDGformer) is proposed. Its core is a dual-driven backbone network that deeply fuses structural and temporal information through alternating G-Blocks and T-Blocks. The backbone is complemented by a parallel synergistic spatio-temporal branch comprising (i) a temporal-driven branch with an attention-guided edge adaptation mechanism to promote information flow and (ii) a structure-driven branch that preserves important structural features to avoid over-globalization. A multi-scale fusion module performs weighted integration of the backbone and branch outputs to enable a comprehensive representation. Moreover, a graph construction method combining k-nearest neighbors (KNN) with random walk encoding is designed to better model latent variable relationships. Experiments on three industrial cases validate that MS-DDGformer achieves superior performance over competing GNN-based, temporal, and GT models.