Domain Knowledge‐Driven Zero‐Shot Learning for Induced Draft Fans Fault Diagnosis

ABSTRACT In this study, a novel domain knowledge‐driven zero‐shot learning framework is proposed for fault diagnosis of induced draft fans in coal‐fired units. By leveraging domain‐specific knowledge and normal data, a CGAN is used to synthesize realistic fault data that closely aligns with actual fault mechanisms. This strategy effectively addresses the challenge of limited fault data, enabling fault diagnosis without relying on historical fault data. Based on the generated fault data, a PCA‐RF‐XGBoost stacked model is developed for multiclass fault classification. Experimental results demonstrate that the proposed method achieves high classification accuracy and robustness across various fault types. Notably, it attains 100% and 96% accuracy in diagnosing motor overload and surge faults, respectively. This study presents a promising zero‐shot learning solution for enhancing the reliability and safety of coal‐fired units under fault data‐scarce conditions.