Casual inference-enabled graph neural networks for generalized fault diagnosis in industrial IoT system

Data-driven fault diagnosis plays a crucial role in diagnosing the operational status within the Industrial Internet of Things (IIoT) systems. Although Graph Neural Networks (GNNs) have recently gained traction in fault diagnosis by adeptly modeling complex dependencies present in high-dimensional sensor measurements, they still grapple with the challenges presented by varying working conditions and pervasive environmental noise, which can significantly hinder their generalization capabilities. Hence, we propose Causal Inference-Enabled Graph Neural Networks (CIE-GNN) for generalized fault diagnosis in large-scale IIoT systems. Specifically, we establish a structural causal model for the GNN-based fault diagnosis model, revealing that the non-causal factors lead to spurious correlations and act as the confounders, thus impairing the generalization performance of fault diagnosis models. To rectify this issue, we design the disentangled transformation module and the causal disentanglement regularization based on mutual information minimization strategy, facilitating the effective decoupling between the fault- causal factors and the non-causal factors in the representation level. Additionally, we propose the random pairing-based backdoor adjustment regularization to mitigate the negative effects of non- causal factors. Extensive experiments and rigorous theoretical analysis validate the generalization capabilities of CIE-GNN across diverse working environments.