Adaptive expert ensembles for fault diagnosis: A graph causal framework addressing distributional shifts

Abstract

Due to the complexity and variability of operating conditions, accurate fault diagnosis is crucial to ensure wind turbine efficiency and reduce maintenance costs. However, existing deep learning models for fault detection often struggle with distribution shifts caused by environmental changes, leading to unreliable predictions. In response, this paper proposes the GCI-ODG framework, an innovative approach based on Graph Causal Intervention (GCI) to enhance Out-Of-Distribution (OOD) generalization in intelligent fault diagnosis of wind turbines. The core innovations of this framework include a hierarchical graph representation that captures both local and global features of multi-condition time-series data, improving the model’s ability to detect intricate fault patterns. An adaptive expert ensemble mechanism is introduced, utilizing pseudo-environment labels inferred without explicit environmental data, enabling dynamic feature extraction and robust adaptation across diverse operating conditions. Additionally, the framework employs causal inference strategies, including backdoor adjustment, to isolate stable, environment-invariant features, effectively mitigating the impact of spurious correlations and distribution shifts. Extensive experiments across multiple benchmark datasets, including real-world wind turbine data, validate the effectiveness of the proposed GCI-ODG framework. The results indicate notable enhancements in both classification accuracy and model robustness under diverse conditions. The GCI-ODG framework demonstrates exceptional capability in handling significant distribution shifts, proving its value as a scalable and generalizable solution for intelligent diagnostics. These findings highlight its potential for reliable and efficient fault detection in complex industrial environments.