Fault Diagnosis using GNN Through Adversarial Attack

The rapid advancement of high-speed trains (HST) has underscored the escalating significance of ensuring the reliability and safety of HST braking systems. In recent years, data-driven fault diagnosis methods have gained traction, owing to the proliferation of sensors employed in fault diagnosis within this domain. Concurrently, fault diagnosis techniques rooted in causal and graph neural networks (GNN) have been gradually integrated into practice. Nevertheless, sensor errors pose significant challenges in the generation of causal graphs. Additionally, the presence of unbalanced data can lead to a deterioration in model performance when classifying minor sample categories. In response, we present an innovative approach for GNN-based fault diagnosis, incorporating adversarial attacks. In this study, we meticulously consider four distinct adversarial attack scenarios, encompassing all potential instances where errors may impact the data. Furthermore, we introduce adversarial regularization to uphold the method's integrity. To substantiate our approach, we conduct fault diagnosis experiments employing graph convolution networks (GCN) and graph attention networks (GAT) models on authentic datasets obtained from HST braking systems. Our empirical findings substantiate the efficacy of our adversarial attack method, which notably enhances the performance of the GNN model in addressing classification issues pertaining to unbalanced fault data.