Causal Intervention-Based GNNs for OOD Generalization in Fault Diagnosis of Wind Turbines

The generalization of machine learning models in out-of-distribution (OOD) scenarios remains a significant challenge, particularly in the context of high-end equipment diagnostics, where dynamic operating environments introduce complex distribution shifts. This study proposes a novel intelligent diagnostic framework based on graph causal intervention, designed to improve model adaptability and robustness under heterogeneous conditions. The framework leverages causal inference principles to infer pseudo-environment labels, enabling the removal of environmental confounding effects without requiring explicit environmental annotations. By integrating causal intervention mechanisms into graph-structured data, the proposed method effectively learns stable causal relationships across diverse environments, enhancing its generalization capabilities. The proposed framework demonstrates reliable performance in addressing OOD generalization challenges, significantly surpassing conventional methods. By dynamically regulating the propagation branch count, it achieves optimal recognition accuracy while reducing redundant computations and feature noise. This study offers a robust and scalable solution for OOD generalization in intelligent diagnostics, providing a foundation for practical applications in high-end industrial systems.

Keywords: Graph causal intervention, Out-of-distribution generalization, Pseudo-environment estimation, Dynamic environmental adaptation, Intelligent diagnostics.