A Multimodal Attention and Hybrid Stacking Framework for Robust Wind Turbine Fault Diagnosis

Abstract

Reliable fault diagnosis in wind turbines is essential for ensuring operational efficiency and preventing costly downtime. This article proposes a novel multimodal attention and stacking framework that integrates domain-driven feature extraction (DDFE), global fusion integration (GFI), and improved attention optimization with hybrid stacking for intelligent fault classification. The framework first extracts discriminative features from mechanical, electrical, and environmental signals using DDFE. These features are then globally fused via GFI to capture intra- and intermodal correlations, followed by an attention-based hybrid stacking ensemble that dynamically weights modalities and combines predictions from multiple classifiers. Extensive experiments on simulated and real-world datasets demonstrate that the framework significantly outperforms conventional single-modality and early-fusion approaches, achieving up to 99.98% accuracy while maintaining robustness under varying operational conditions. The results highlight the advantages of combining physically interpretable features with data-driven multimodal learning for highly reliable wind turbine fault diagnosis.