Abstract: [Objective] The operation and maintenance (O&M) paradigm of wind power stations is undergoing a profound transition from corrective repair after failure to predictive maintenance before failure. The emergence of industrial artificial intelligence (AI) foundation models opens a new technical pathway for this transition, yet their applicability boundaries, methodological system, and engineering constraints in the wind energy sector remain unsystematically examined. Following a data–model–application–decision chain, this paper systematically surveys and synthesizes representative literature since 2009 on condition monitoring, fault diagnosis, and remaining useful life prediction for wind turbines. It critically distinguishes the mechanisms and applicability boundaries of four technical routes—physics-based models, conventional machine learning, deep learning, and industrial AI foundation models—and analyzes engineering outcomes across typical targets (gearbox, main bearing, blade, pitch system) and station-level O&M decision scenarios. [Results] The findings show that: (1) wind turbine O&M technology exhibits a four-stage evolution from threshold alarming through condition identification and life prediction to autonomous decision-making, in which the fusion of SCADA and condition monitoring system data is a prerequisite for methodological effectiveness; (2) deep learning has markedly surpassed conventional methods in component-level diagnostic accuracy, yet engineering reliance remains lower than expected owing to label scarcity, time-varying operating conditions, and insufficient cross-turbine generalization; (3) industrial AI foundation models centered on time-series foundation models and large language models demonstrate clear advantages in few-shot, zero-shot, and cross-turbine transfer settings through large-scale pre-training, cross-modal reprogramming, and domain knowledge injection, and have begun to form preliminary engineering loops in fault reasoning and maintenance strategy generation; and (4) deployment still faces prominent obstacles including data silos, hallucination risk, real-time constraints, insufficient interpretability, and the absence of evaluation standards. [Conclusions] To achieve proactive prevention in wind power stations, future efforts should prioritize domain-specific foundation models for wind energy, physics–data fused world models, edge–cloud collaborative autonomous O&M agents, and quantifiable and trustworthy evaluation benchmarks, thereby advancing predictive maintenance from isolated algorithmic breakthroughs toward systemic capability building.
