Local Feature Augmented Graph Transformer for Enhanced Industrial Process Fault Diagnosis

Industrial fault diagnosis is a core technical support for ensuring the safe and efficient operation of modern industrial measurement systems. However, the key challenges of complex spatiotemporal correlations and sparse local fault features in sensor network data severely constrain the measurement accuracy of traditional diagnostic methods. To address these challenges, this article proposes a novel local feature augmented graph transformer (LFA-GTF) framework, which integrates global topology modeling and local signal enhancement techniques. First, the framework constructs a graph transformer (GTF) architecture, which accurately captures global spatiotemporal dependencies in industrial data by integrating a multihead self-attention mechanism (SAM). Second, a local feature augmentation (LFA) module is designed to mitigate feature sparsity by adaptively enhancing discriminative local patterns. Comprehensive experiments on the revised Tennessee Eastman (TE) dataset demonstrate that the LFA-GTF framework achieves a fault classification accuracy of 96.06%, outperforming baseline graph neural networks (GNNs) such as graph convolutional network (GCN), graph attention network (GAT), and graph sample and aggregate network (GraphSage) by 14.3%–41.7%. Ablation studies verify that the LFA module improves classification performance by 5.0%–13.7% across fault types with sparse features, and it also exhibits good generalization ability on three-phase flow (TPF) processes. This framework provides a powerful solution for fault diagnosis of industrial measurement systems under unbalanced data conditions.